<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[SISSA mathLab - Medium]]></title>
        <description><![CDATA[SISSA mathlab is a laboratory for mathematical modeling and scientific computing devoted to the interactions between mathematics and its applications. - Medium]]></description>
        <link>https://medium.com/sissa-mathlab?source=rss----2639606b2015---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>SISSA mathLab - Medium</title>
            <link>https://medium.com/sissa-mathlab?source=rss----2639606b2015---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 22:39:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/sissa-mathlab" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Introduction to Fractional Calculus]]></title>
            <link>https://medium.com/sissa-mathlab/introduction-to-fractional-calculus-4b62f1ff094a?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/4b62f1ff094a</guid>
            <dc:creator><![CDATA[Pavan Pranjivan Mehta]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 15:16:33 GMT</pubDate>
            <atom:updated>2026-07-12T15:16:33.168Z</atom:updated>
            <content:encoded><![CDATA[<p>Fractional Calculus is a generalization of integer-order calculus that extends the order differential and integral from the natural numbers ℕ to the positive real numbers ℝ+. Broadly speaking, there are two well-known approaches to such an extension.</p><p>(1) Starting from the limit definition of a derivative, we obtain the Grünwald-Letnikov fractional derivative [1-3].</p><p>(2) Generalizing Cauchy’s formula for repeated integration yields the Riemann-Liouville fractional integral, which is subsequently used to define the Riemann-Liouville and Caputo fractional derivatives [2-3].</p><p>Paradoxically, a fractional derivative is no longer a local operator, but a non-local one. The notion of Brownian motion is extended to admit Lévy stable processes in the case of fractional diffusion, thus making them appropriate mathematical objects to address non-local phenomena.</p><p>In the case of turbulence [4, 5] the duality of local and non-local processes in turbulence is modeled with a variable order Caputo fractional derivative. Furthermore, fractional and non-local operators have found real-world applications in fracture mechanics (peridynamics) [6, 7], economics and finance [8, 9], viscoelasticity, electrical circuits, and plasma physics.</p><p>In the following section, we introduce the definitions of fractional derivatives and integrals.</p><h3>Grunwald-Letnikov Definition</h3><p>To derive the Grünwald-Letnikov fractional derivative [1] (see also [2]), consider successive backward dfferences:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*KCBVTHAhzlKdfRZqf5LZZQ.png" /><figcaption>Eq. 1</figcaption></figure><p>Therefore, for a <em>p-</em>th order differentiation, where <em>p</em> ∈ ℕ (and <em>f</em> ∈ <em>C</em>[<em>a</em>, <em>b</em>]), we obtain:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/393/1*k90ALi3P-q_LtphnHzNQIg.png" /><figcaption>Eq. 2</figcaption></figure><p>where the binomial coefficients are defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/367/1*ecrfCgsERmLRs3JRLBzvkA.png" /><figcaption>Eq. 3</figcaption></figure><p>Equation (2), originally derived for <em>p</em> ∈ ℕ, can be equivalently rewritten by constructing a uniform grid on <em>x</em> ∈ [<em>a</em>, <em>b</em>], with N = (<em>x </em>− <em>a</em>)/<em>h</em>. Although this formulation is derived for <em>p</em> ∈ ℕ, it is valid for any non-integer positive order (<em>p</em> &gt; 0). Thus, the Grünwald-Letnikov fractional derivative is defined for all <em>p</em> ∈ ℝ+. However, for non-integer orders, the terms in the infinite series do not vanish. Consequently, Diethelm [3] imposed an additional condition on the function <em>f</em>, given by the zero-extension in Equation (4); its equivalence is argued in Theorem 2.25 of [3]:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/1*BTqXyrO-hWcQhoKLSb15ig.png" /><figcaption>Eq. 4</figcaption></figure><p>To extend this definition to a negative order , we substitute <em>-p</em> into the formulation, yielding:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/430/1*yT0lIBwmc7oJTFktfym6tg.png" /><figcaption>Eq. 5</figcaption></figure><p>where the negative binomial coefficient is expanded as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/1*Kly3bo3ny6Dh1_McB596vw.png" /><figcaption>Eq. 6</figcaption></figure><p>Following [2], this notation implies that:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/364/1*NxWHkiXiLBA_jgAv_VKpkA.png" /><figcaption>Eq. 7</figcaption></figure><p>Thus, the Grünwald-Letnikov fractional integral is defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/356/1*8BcCiw8RYBw8XH2B0vnVCA.png" /><figcaption>Eq. 8</figcaption></figure><p><strong>Remark 1.</strong> Equations (2) and (8) provide a unified framework for the fractional derivative and fractional integration, respectively, for arbitrary order <em>p</em> ∈ ℝ.</p><h3>Riemann-Liouville Definition</h3><p>In this section, we present the definition of the Riemann-Liouville operators. Recall Cauchy’s formula for repeated integration for <em>p</em> ∈ ℕ:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*UCQ2tI4AHhcldsg5LTBZhw.png" /><figcaption>Eq. 9</figcaption></figure><p>Here, for <em>p</em> ∈ ℕ, we can generalise this relationship by introducing the Euler gamma function Γ(<em>p</em>) = (<em>p</em> − 1)!. Consequently, the Riemann-Liouville fractional integral is defined for any <em>p</em> ∈ ℝ+ as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/334/1*1tdUKmt5ydoQG4MM-Yy3vQ.png" /><figcaption>Eq. 10</figcaption></figure><p>Note the notation used for the Riemann-Liouville fractional integral, which denotes a fractional integration of order <em>p </em>with a lower limit <em>a</em> and an upper limit <em>x</em>. Additionally, Theorem 2.1 of [3] provides the proof of its existence:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/567/1*HivBy1e3p3_3osSQL4B_XQ.png" /><figcaption>Eq. 12</figcaption></figure><p>Furthermore, Theorem 2.2 of [3] establishes the semigroup property of Riemann-Liouville fractional integrals for <em>p</em>, <em>q</em> ∈ ℝ+:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/233/1*aDgXdnavDelEu1gAsl6QfA.png" /><figcaption>Eq. 13</figcaption></figure><p>Subsequently, the Riemann-Liouville fractional derivative of order <em>p</em> is defined for <em>p</em> ∈ ℝ+ and <em>k</em> ∈ ℕ:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/567/1*3DhTTyyJNc1-QhY5Wh3pNQ.png" /><figcaption>Eq. 14</figcaption></figure><p>This allows the Riemann-Liouville fractional derivative to be compactly expressed in terms of the Riemann-Liouville fractional integral as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/1*t0-IuHCulOy7TBbxDoqH3Q.png" /><figcaption>Eq. 15</figcaption></figure><p>Lemma 2.12 of [3] provides the existence proof for the Riemann-Liouville fractional derivative. Indeed, for <em>p</em> ∈ ℕ, the Riemann-Liouville fractional derivative reduces to the classical integer-order derivative. Furthermore, Theorem 2.13 of [3] establishes an identity for the composition of these operators for <em>p</em>, <em>q</em> ∈ ℝ+</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/210/1*2A0rfCfyKiMiTbvIOcTrzg.png" /><figcaption>Eq. 16</figcaption></figure><p><strong>Remark 2.</strong> Upon performing the limit Equation (2), it can be shown that the Grünwald-Letnikov fractional derivative coincides with Riemann-Liouville fractional derivative. Thus the two definitions are equivalent (see Section 2.3.7 of [2] and Section 3.3 of [25]).</p><h3>Caputo Definition</h3><p>Although, the Riemann-Liouville fractional derivative is mathematically well established, it presents two key problems when applied to physical systems (see Section 2.4 of [2]).</p><ul><li>The Riemann–Liouville fractional derivative of a constant function is, in general, non-zero. It vanishes only in the special case where <em>a</em> → −∞. This behavior is inconvenient for modeling physical systems.</li><li>The required initial conditions take the form limit <em>t</em>→<em>a</em> <em>RLa</em> <em>D</em>^<em>p</em>_<em>t</em> <em>f</em> (<em>t</em>) = <em>g</em>. Such initial conditions lack physical intuition,whereas real-world applications typically require explicitly measurable function values such as <em>f</em>(<em>t</em>)|(<em>t</em>=0) = <em>g.</em> (See Section 2.4 of [2] for a detailed elaboration on how these requirements complicate the application of the the Laplace transform).</li></ul><p>To addres this, Caputo defined a new fractional derivative [26] as (15) for <em>p</em> ∈ ℝ+ and <em>k</em> ∈ ℕ, as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*8kZQ2J9s6PNy3wZu_CiGlA.png" /><figcaption>Eq. 15</figcaption></figure><p>Indeed, the Caputo fractional derivative of a constant function is zero, and it associated initial conditions can be formulated in a classical way. Theorem 3.1 of [3] expresses the Caputo fractional derivative in terms of Riemann-Liouville fractional derivative by introducing a Taylor polynomial.</p><p>Furthermore, Theorem 3.7 of [3] establishes the following fundamental inversion property:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/197/1*rs0t_0l1NKpe5L4cci_J6w.png" /><figcaption>Eq. 16</figcaption></figure><p><strong>Remark 3.</strong> For homogeneous initial conditions, the Caputo fractional derivative is equivalent to the Riemann-Liouville fractional derivative. This identity can be shown by applying intergation by parts repeatedly.</p><p>The semigroup property of the Caputo fractional derivative is expressed as follows for <em>p</em>, <em>q</em> &gt; 0, which holds when <em>p</em> and <em>q</em> are integers as well (see Lemma 3.13 of [3]):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/274/1*PxfC17ohhVATUoiDetkwRQ.png" /><figcaption>Eq. 17</figcaption></figure><h3>Left and right sided fractional derivative</h3><p>Upon closer inspection of the definitions introduced in the preceding sections, it is evident that these fractional derivatives only accounts for non-local interactions from the left; hence, they are termed left fractional derivatives. Conversely, a right fractional derivative can also be defined to account for non-local interactions from the right. Incorporating both perspectives ensures that the mathematical theory of fractional calculus is complete. In this post, we employ both the left and right definitions, as physical processes represent an integration of interactions from both the left and right boundaries.</p><p>For clarity, we restate the left Riemann-Liouville fractional derivative for a function defined over an interval [<em>a</em>, <em>b</em>], <em>p</em> ∈ ℝ+ and <em>k</em> ∈ ℕ:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/566/1*L_E4a3oXWaYlkOK3UATxpg.png" /><figcaption>Eq. 18</figcaption></figure><p>Conversely, the right Riemann-Liouville fractional derivative is defined for <em>p</em> ∈ ℝ+ and <em>k</em> ∈ ℕ:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/567/1*uQI_lyYliWZxOStmT8FH5A.png" /><figcaption>Eq. 19</figcaption></figure><p>Note that the domain of integration spans from <em>x</em> to the upper bound <em>b</em>.<br>Similarly, the left Caputo fractional derivative for a function defined over an interval [<em>a</em>, <em>b</em>], <em>p</em> ∈ ℝ+ and <em>k</em> ∈ ℕ is given by:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/563/1*h6ZGIeT0Fz1ClC_8Y3opcw.png" /><figcaption>Eq. 20</figcaption></figure><p>Meanwhile, the right Caputo fractional derivative is defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/558/1*eYmVFKpFrKUKwzngvsYhSg.png" /><figcaption>Eq. 21</figcaption></figure><p>where, once again, the integration limits are inverted.<br>The combination of these left and right definitions allows us to explicitly define the <strong>Riesz fractional derivative </strong>for a function defined over an interval [<em>a</em>, <em>b</em>], with <em>c_p</em> = −0.5<em>cos</em>(<em>p</em>π/2):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/382/1*P9sy6T3nnbmDu1ynGvRF_w.png" /><figcaption>Eq. 22</figcaption></figure><h3>References</h3><p>[1] Grunwald AK. ¨Uber ”begrente” derivationen und deren anwedung. Zangew Math und Phys. 1867;12:441–480.</p><p>[2] Podlubny I. Fractional differential equations. Academic; 1999.</p><p>[3] Diethelm K. The analysis of fractional differential equations, volume 2004 of lecture notes in mathematics. Springer-Verlag, Berlin; 2010.</p><p>[4] Pavan P. Mehta, Guofei Pang, Fangying Song, and George Karniadakis. Discovering a universal variable-order fractional model for turbulent couette flow using a physics-informed neural network. Fractional Calculus and Applied Analysis, 22(6):1675–1688, 2019.</p><p>[5] Pavan Pranjivan Mehta. Fractional and tempered fractional models for reynolds-averaged navier–stokes equations. Journal of Turbulence, 24(11–12):507–553, 2023.</p><p>[6] Madenci, E., Anicode, S.V.K. and Zhang, Y., 2024, Peridynamics and Its Applications Using Ansys, <em>Springer</em>, Boston, MA.</p><p>[7] Madenci, E., Roy, P. and Behera, D., 2022, Advances in Peridynamics, <em>Springer</em>, Boston, MA.</p><p>[8] Scalas, E., Gorenflo, R., &amp; Mainardi, F. (2000). Fractional calculus and continuous-time finance. <em>Physica A: Statistical Mechanics and its Applications</em>, <em>284</em>(1–4), 376–384.</p><p>[9] Mainardi, F., Raberto, M., Gorenflo, R., &amp; Scalas, E. (2000). Fractional calculus and continuous-time finance II: the waiting-time distribution. <em>Physica A: Statistical Mechanics and its Applications</em>, <em>287</em>(3–4), 468–481.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4b62f1ff094a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/introduction-to-fractional-calculus-4b62f1ff094a">Introduction to Fractional Calculus</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Comparative study of fatigue damage modeling approaches]]></title>
            <link>https://medium.com/sissa-mathlab/comparative-study-of-fatigue-damage-modeling-approaches-bbcbf9adffd8?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/bbcbf9adffd8</guid>
            <dc:creator><![CDATA[Elvin Babayev]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 14:51:56 GMT</pubDate>
            <atom:updated>2026-07-12T14:51:55.080Z</atom:updated>
            <content:encoded><![CDATA[<p>In the modern industrial landscape, specifically in heavy machinery, automotive and motor sector, rolling-element bearing failures are the most common reason of failure of such machines; they represent about 40–50 % of every failure in any industrial arrangement (Lei et al., 2018). Mainly because of the fatigue models that consider surface pitting damage, spalling damage, and crack initiation, there has been an increasing need to use the Prognostics and Health Management (PHM) methods to predict these failures by calculating the Remaining Useful Life (RUL) (Wang et al., 2024). Therefore, the shift from reactive to proactive maintenance frameworks is no longer a matter of choice but a critical requirement for operational resilience.</p><p>Effective Predictive Maintenance (PdM) activities in rotating machines largely depend on accurate estimation of the Remaining Useful Life (RUL) of parts in rotating machines. On this study we presents two fatigue models that were developed to estimate the wear of bearing rollers. The models are built and assesseed with help of the NASA IMS benchmark dataset that is widely regarded. <strong>Method 1</strong>. uses a hybrid analytical model by combining the cumulative damage rule of Miner with a neural network (NN) to predict stress caused by features based on vibration. In comparison, <strong>Method 2.</strong> maintains a physics-based model of failure which is concerned with transforming raw vibration data directly into stress history through counting cycles to assess the accumulation of fatigue damage. Both models modify Basquin parameters to bring theoretical failure times to meet empirical observed results. Empirical findings prove that <strong>Method 2. </strong>provides the data with improved continuity and directness in terms of Degradation Energy Indicator (DEI) and is closer to the real failure time.</p><p>The NASA IMS dataset contains no direct data about material properties or geometry or applied loads which makes analytical stress computation impossible. We used a data-based method to train NN which produces synthetic stress signals that follow physical laws based on vibration measurement data. The NN model detects the intricate non-linear patterns which exist between statistical vibration features and their corresponding stress responses. The input feature set contains normalized values that include RMS, crest factor, kurtosis, clearance factor, shape factor, skewness, and standard deviation which represent different characteristics of vibration energy distribution and dynamic response.</p><p>The target synthetic stress function can be expressed as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*LHTCebCEjAArXf_-T4gfcw.png" /></figure><p>Where, 𝜎 target is the synthetic stress signal to be approximated,</p><p>𝑋 features 𝑖 denotes the normalized statistical features,</p><p>𝑐<em>i</em> are scaling coefficients representing feature weights,</p><p>𝑓𝑖(⋅) are nonlinear transformation functions modeling individual degradation mechanisms</p><p>𝐼𝑗 (𝑋features) represents interaction terms capturing coupled effects of degradation.</p><p>After successfully generating synthetic stress from extracted features, we proceed with damage accumulation at each life cycle of bearing. These cycles represent the alternating load reversals that contribute to fatigue degradation within the bearing material. Subsequently, the total accumulated cumulative damage (<em>D</em>) is calculated using the linear damage accumulation hypothesis, more widely known as Miner’s Rule, which is expressed as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/272/1*H5hFTukDN6Wk4v4NWiyfgg.png" /></figure><p>Where 𝐷 is the total cumulative damage, 𝑘 is the number of distinct stress levels 𝜎𝑖, 𝑛𝑖 is the number of applied cycles at each stress level, and Ni represents the number of cycles to failure corresponding to that stress amplitude. Once the cumulative damage 𝐷(𝑡) is obtained, it is transformed into DEI and it provides a normalized measure of degradation that decreases monotonically from unity (representing a fully healthy condition) toward zero (indicating complete failure). It is defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/221/1*eKYRtIP5ZabwQbluS1sdzA.png" /></figure><p>The DEI curve serves as the primary prognostic metric in this framework, reflecting the progressive loss of structural integrity over time. By tracking its temporal evolution, the RUL can be estimated through extrapolation of the DEI trend toward the defined failure threshold (Fig 1.3.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*TCt8gyDp4O4FpTs4IDVCqQ.png" /></figure><p>As observed in Figure 1.3, the DEI shows abrupt drops corresponding to transient spikes in the vibration signal, rather than a smooth and gradual decline. This behavior is associated with the first bearing in the dataset, which was identified as faulty by NASA, with outer race failure occurring near the end of the test sequence. Although the extracted statistical features collectively provide valuable information for RUL estimation, their effective application requires careful feature engineering and parameter tuning to ensure stable and physically consistent degradation trends. Results for 3rd and 4th bearing is similar to bearing 1 at this stage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/964/1*jo3QfqZ882-5prU1T2-u9g.png" /></figure><p>In method 2 instead we consider accelerometer data as direct stress function, even though they are different in nature, there’s a strong correlation between them as many other studies confirm. Before the vibration signal can be used for fatigue analysis, it must be processed carefully to eliminate noise, low-frequency drift, and high-frequency components that do not relate to material stress.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*zfuFo0bm83o-omA0608JXA.png" /></figure><p>For this method cycle extraction has been done using rain-flow counting rather than amplitude counting in the first method, since rainflow algorithm gives broad and detailed info of cycles, and again Miner’s rule was applied to calculate the damage and finally the degradation index. If you pay close attention to Figure2.3, there’s more stable and clear understanding of machine failure life starting from its healthy time until It dies, and the graph perfectly illustrates it. This monotonic behavior makes 𝐷(𝑡) a reliable quantitative indicator of the bearing’s health evolution. By analyzing the slope and curvature of this trend, one can identify the transition from the normal operating regime to the critical damage phase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*lDURLqLrdKmUOI6DnsoJFg.png" /></figure><p>As conclusion we can say that the comparison of the two methods showed that both can predict the wear behavior of bearings. However, Method II produced more consistent and realistic results. Its steps for filtering, integration, and noise removal effectively isolated the true stress response from the raw vibration data. This allowed the Rain-flow algorithm and Miner’s Rule to work on a clean and meaningful signal. As a result, it led to smoother degradation trends, greater consistency in the Degradation Energy Indicator, and better reliability in estimating Remaining Useful Life. Overall, using fatigue algorithms directly on the processed stress data in the second method provided more detailed and understandable results. This confirmed its advantage for practical condition monitoring and forecasting.</p><p><strong>References</strong></p><p>1. Lee, J., Qiu, H., Yu, G., Lin, J., and Rexnord Technical Services. (2007). <em>Bearing Data Set. </em>NASA Ames Prognostics Data Repository, NASA Ames Research Center,<em> </em>Moffett Field, CA.</p><p>2. Wang, P., Chen, M., &amp; Yang, H. (2024). <em>A Comprehensive Review of Remaining Useful Life Estimation Approaches for Rotating Machinery. </em>Energies, 17(22), 5538.</p><p>3. Schaeffler Group. (2017). <em>The Role of Vibration Monitoring in Predictive Maintenance. </em>Technical Paper.</p><p>4. Miner, M. A. (1945). <em>Cumulative Damage in Fatigue. </em>Journal of Applied Mechanics, 12(3), A159–A164.</p><p>5. Lei, Y., Lin, J., Zuo, M. J., &amp; He, Z. (2017). <em>A Review on Machinery Prognosis: State of the Art and Challenges. </em>Journal of Mechanical Science and Technology, 31, 3501–<em> </em>3518.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbcbf9adffd8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/comparative-study-of-fatigue-damage-modeling-approaches-bbcbf9adffd8">Comparative study of fatigue damage modeling approaches</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Physics Informed Neural Networks applied to Conservation Laws with discontinuities]]></title>
            <link>https://medium.com/sissa-mathlab/physics-informed-neural-networks-applied-to-conservation-laws-with-discontinuities-23303b3a8106?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/23303b3a8106</guid>
            <dc:creator><![CDATA[Guglielmo Padula]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 14:51:17 GMT</pubDate>
            <atom:updated>2026-07-12T14:51:15.913Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>In this post, we explore how <strong>Physics-Informed Neural Networks (PINNs)</strong> — a class of neural networks designed to solve <strong>Partial Differential Equations (PDEs)</strong> — can be applied to <strong>conservation laws with discontinuities</strong>.</p><p>We begin with a brief overview of how PINNs work, followed by an introduction to conservation laws and the challenges posed by discontinuities. Finally, we present a series of experiments aimed at gaining insights into how PINNs can be effectively applied to these problems.</p><h3>Physics Informed Neural Networks</h3><p>Let us suppose that we have a differential equation of the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jk_Gj60anL0Uv5CK-iCbXw.png" /></figure><p>where <em>B</em> and <em>F</em> may contain derivative terms, like <em>du/dx</em>. Given a set of points</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zTWuOzRO11p10dpPyv8Jdg.png" /></figure><p>at which we want to evaluate the solution, one approach is to minimize:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QT0jYokc-XbOjqXLLY-fsg.png" /></figure><p>over a given function class <em>A</em>. If the points <em>x_{i}</em> coincide with the Gauss-Lobatto nodes in 1D, and <em>A</em> is the function class of polynomials of degree at most <em>N</em>-1, the problem reduces to a spectral method. In this post, we focus on the class of neural networks with the following architecture:</p><ul><li>depth: 5 layers,</li><li>width: 5 neurons per layer,</li><li>input dimension:2,</li><li>ouput dimension: 1,</li><li>ReLu activation function,</li><li>a modulo-1 activation layer preceding the first linear layer, to enforce periodicity.</li></ul><h3>Scalar conservation Law</h3><p>Conservation laws are PDEs of the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xkI1KyEL6NjQwp4MYJorDQ.png" /></figure><p>We restrict the problem to an initial condition in time and a periodic boundary condition in space, yielding the system:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1-qxoTIxLxUe1kFHIoc8uw.png" /></figure><p>Note that even with a smooth initial condition, a classical solution that satisfies the equation everywhere globally in time and space may not exist. Therefore, we consider the weak form instead of the strong form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vlXrrYH36lihU3UywGplYA.png" /></figure><p>A solution to teh weak formulation is guaranteed to exist; however, it may not be unique. Nevertheless, it can be proven that there is a unique, physically admissible solution, namely the one obtained as the limit:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*yeohThUCNinuoPAaqc1FJg.png" /></figure><p>where <em>u_{epsilon}</em> is the unique solution to the regularized system:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UcKOfPKX9w-Th8014E3nhw.png" /></figure><p>This particular solution, <em>u</em>, is also the only one that satisfies the following entropy inequality:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fkTIubowuVb4cUke_8xTTg.png" /></figure><p>called the Kruzkov entropy condition. For this reason, we call this particular solution the entropy solution.</p><p>We can now turn to the problem that we investigate in this blog post.</p><h3>Can we approximate entropy solutions of Conservation Laws using Physics Informed Neural Networks?</h3><p>To address this question, lets first study the easiest form of scalar conservation law, the transport equation, which is obtained with <em>F(u) = u</em>. As we don’t want the problem to be too easy, we adopt a discontinous initial solution. The system that we consider is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DQmu-cudMurMH0ji9PIw-A.png" /></figure><p>with <em>T</em> = 1/8 and</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p7eHeoWUvb_LoXMO3AzJQg.png" /></figure><p>This system has an explicit solution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aOEpZS-tM3gbDOZ28p2Uxw.png" /></figure><p>First, we try to learn the solution by adopting a classic PINN loss:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gv38hsK7yeakoN_txuMDQw.png" /></figure><p>where the integrals are computed with a uniform grid in space and time, and in a Monte Carlo manner, and <em>u_{theta}</em> is our reference neural network. Let’s see the plot of <em>u_{theta}</em> after parameter optimization compared with the true solution at time <em>T</em> = 1/8, along with the train loss. The periodic boundary condition is not imposed because it is automatically satisfied due to the neural network structure. The optimization is performed using the PINA library [1].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*nkMIH_DaZYcqDPC3cy4xJw.png" /><figcaption>Fig. 1: Classical PINN result for the transport equation.</figcaption></figure><p>We can observe that the neural network approximates the true solution pretty well and that the training loss decreases towards 0. The fact that PINNS well approximate linear problems is well known in the literature [1, 2], so we do not discuss this. However, one may ask himself why the neural network converges to the entropy solution, and not to another weak solution. This is due to the implicit regularization of neural network, which induces an artificial diffusion in the PDE loss, and thus the neural network converges to the entropy solution as the loss goes to 0.</p><p>Let’s try to understand if we are able to generalize this behavior, for example by adopting a hard PINN loss, so by incorporating the boundary condition in the model. The system now becomes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tB7rRknpy9xOozj24k_6qw.png" /></figure><p>Let’s see how well this model behaves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*tdPSDqwpYH5o82pyhKyMzA.png" /><figcaption>Fig. 2: Hard PINN result for the transport equation.</figcaption></figure><p>An interesting phenomenon occurs: the training loss goes to 0, while the PINN solution remains stuck at the initial condition. This happens because the initial condition is itself a weak solution, but not an entropy solution; thus, the trivial neural network (the constant zero solution) is preferred. Another way to see this is that no artificial diffusion is introduced because the model is not smooth (since the initial condition <em>u_0</em> is not even continous). Therefore, the convergence of the loss to 0 does not guarantee convergence to the entropy solution.</p><p>Now we can see how these results generalize when <em>F</em> is no longer the identity, meaning the PDE is not longer linear. In this case, the system may develop shocks (in the sense that the gradient can become infinite).</p><p>Let’s study the PDE known as the Burgers’ equation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aq_GvUgX6ACUyqReqbeWSA.png" /></figure><p>with the same initial condition as before. Advantageously, the exact solution remains identical to the previous one. This choice ensures that the same neural network architecture used for the transport equation can also be applied here without losing representation power. <br>Next, let us adopt the classic PINN loss function:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5wOwCH-JsgrzrAGJ7Wa8Ew.png" /></figure><p>and let’s see how the neural network training behaves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*Wna7YE1djv33Nc_SjuUwEw.png" /><figcaption>Fig. 3: Classical PINN result for the Burgers’ equation.</figcaption></figure><p>The loss converges to a non-global minimum, resulting in a very poor solution. This occurs because pure PINNS are not, in general, able to learn nonlinear scalar conservation laws [2] due to the intrinsic structure of the PDE.</p><p>To address this issue, one approach is to introduce data-driven regularization. We assume the existence of a binary function, <em>w(t, x)</em>, with a finite number of jumps, which indicates the points at which the true function is known. The loss function then becomes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ryMAdg6c411VzRN7Zmpnhw.png" /></figure><p>We can now examine the results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*-wNLA0H_p1x-Gc90AgjnUQ.png" /><figcaption>Fig. 4: Hybrid PINN results for the Burgers’ equation.</figcaption></figure><p>The training loss converges to 0, and the PINN solution accurately approximates the true solution.</p><h3>Conclusion</h3><p>We have seen how to make Physics Informed Neural Networks work with conservation laws. With linear conservation laws, no additional modifications are needed, but with non-linear conservation laws, classic PINNs are not enough. An alternative is to add a loss term based on the Kruzkov entropy condition [2,4]. The code for the figures can be found <a href="https://github.com/guglielmopadula/DiscConsLawMedium">here</a>.</p><h3>References</h3><p>[1] PINA, <a href="https://github.com/mathLab/PINA">https://github.com/mathLab/PINA</a>.</p><p>[2] Improving Weak PINNs for Hyperbolic Conservation Laws: Dual Norm Computation, Boundary Conditions and Systems, A. Chaumet, J. Giesselmann.</p><p>[3] On the convergence of PINNs. N. Doumeché, G. Biau, C. Boyer.</p><p>[4] wPINNs: Weak Physics informed neural networks for approximating entropy solutions of hyperbolic conservation law. T. D. Ryck, S. Mishra, R. Molinaro.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23303b3a8106" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/physics-informed-neural-networks-applied-to-conservation-laws-with-discontinuities-23303b3a8106">Physics Informed Neural Networks applied to Conservation Laws with discontinuities</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Embedding Machine Learning with Computational Fluid Dynamics simulations: a possible marriage?]]></title>
            <link>https://medium.com/sissa-mathlab/embedding-machine-learning-with-computational-fluid-dynamics-simulations-a-possible-marriage-43ee738c6c4e?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/43ee738c6c4e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fluid-dynamics]]></category>
            <category><![CDATA[computational-physics]]></category>
            <category><![CDATA[turbulence-modeling]]></category>
            <dc:creator><![CDATA[Nico clinco]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 14:49:37 GMT</pubDate>
            <atom:updated>2026-07-12T14:49:35.933Z</atom:updated>
            <content:encoded><![CDATA[<p>Computational Fluid Dynamics (CFD) is essential for turbulence modeling as it provides an efficient framework for solving complex equations governing turbulent flows. Significant advancements in computational resources have enhanced CFD’s capability to simulate increasingly complex turbulent flows. In this scenario, the most common approaches for turbulence modeling can be classified into three big blocks, depending on the level of accuracy we wish to obtain:</p><ol><li>Direct Numerical Simulation (DNS): This approach is the most accurate but the most expensive one. It can be seen as the “Business class” of the simulations.</li><li>Reynolds Averaged Navier-Stokes Equations (RANS): This approach does not require big computational resources and corresponds to the equivalent “Economy class”.</li><li>Large-Eddy Simulation (LES): This approach would be a trade-off between the DNS and the RANS approach. It allows us to achieve a good trade-off between accuracy and computational cost.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/441/1*xJhpTq1X9A55dTSiFENY5g.png" /><figcaption>Fig.1: Three different approaches adopted for turbulence modeling.</figcaption></figure><p>For the majority of cases of interest, a DNS is unfeasible, as it requires a really high computational power. Consequently, we most often rely on the other two approaches, namely the LES or RANS.</p><p>However, these two approaches, namely, the RANS or LES require modeling. More specifically, the various models proposed in these frameworks are affected by many parameters, and consequently, they are affected by uncertainty.</p><p>Recently, with the widespread use of machine learning, data-driven strategies have been proposed in order to mitigate this problem [1]. Briefly speaking, we aim to correct deficiencies of various models by leveraging high-fidelity data (generated most often by a DNS). A schematic representation is presented in Fig. 2, illustrating how high-fidelity data, interpolated onto the computational grid of the target simulation, are employed to train a model parameterized by a set of variables θ. The resulting model is then used as a corrective term within the simulation framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/558/1*4JUaSazgrdsJdQPgC-4X_A.png" /><figcaption>Fig. 2: A schematic drawing of the framework: high-fidelity data are used for training a model which depends by many parameters.</figcaption></figure><p>The usage of Machine Learning for turbulence modeling not only allows us to create ad-hoc models with higher accuracy, but also to create models which depend directly upon specific flow features, extracted directly from high fidelity data. These features can be extracted directly from high-fidelity data, such as DNS, allowing the learned models to capture complex, nonlinear interactions that are often difficult to capture using traditional closure assumptions. As a result, machine-learning-based turbulence models offer a data-driven framework that can adapt to different flow regimes, improve generalization across configurations, and provide deeper insight into the underlying physics governing turbulent flows.</p><h4><strong>Offline vs Online training in the context of turbulence modeling</strong></h4><p>In the context of data-driven turbulence modeling the training can be done into two different ways, namely in the offline or online fashion (so-called end-to end). Offline training relies on pre-collected high-fidelity datasets, and then uses this data to train a model the will be incorporated in the solver only after the whole training <strong>[1]</strong>. This approach allows for extensive training and validation at relatively low computational cost during inference, but its performance may degrade when applied to flow conditions that deviate significantly from the training data. A schematic view is shown in Fig.3. In this scenario, both high-fidelity and low fidelity snapshots are collected a-priori. They are used in the training of the model in order to find the best parameters of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/906/1*AaAK0hfrQ2FQ_WnYncAJjg.png" /><figcaption>Fig. 3: A schematic view of the <strong>offline training</strong> approach. Here HF and LF are referred as high-fidelity and low fidelity data.</figcaption></figure><p>In contrast, online training (see Fig.4) updates the model during the solution process and thus adjust the model parameters depending on the evolving flow conditions <strong>[1,2,3]</strong>. While this approach can improve robustness and accuracy in complex flows, it introduces additional computational overhead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*SsiBKlQmwxCDG_0VcG0Ybw.png" /><figcaption>Fig. 4: A schematic view of the <strong>online training</strong> approach. The solver is used for training our model.</figcaption></figure><p>In general, the second approach is more effective; however, for complex simulations with a large number of degrees of freedom, it remains computationally unfeasible. To address this limitation, some researchers have proposed intermediate strategies that aim to balance the advantages of both approaches. For instance, in <strong>[3]</strong>, the training data are generated by restarting coarse simulations from high-fidelity snapshots, thereby enriching the dataset while limiting computational cost. Although this methodology has been shown to be effective in certain scenarios, it may not provide a universally optimal solution across different flow configurations.</p><p>In conclusion, with the increasing availability of high-fidelity simulation data and growing computational resources, data-driven methods are emerging as powerful tools for turbulence modeling. Within this framework, models can be trained either in a purely offline manner or updated online within the flow solver. While online training generally offers improved robustness and adaptability, it comes at a higher computational cost compared to offline approaches, highlighting an inherent trade-off between accuracy and efficiency.</p><h4>References</h4><ol><li>Karthik Duraisamy, Gianluca Iaccarino and Heng Xiao. “<em>Turbulence Modeling in the Age of Data</em>”. Annual Review of Fluid Mechanics 51(2019): 357–377.</li><li>Deniz A. Bezgin , Aaron B. Buhendwa, Steffen J. Schmidt, Nikolaus A. Adams. “<em>ML-ILES: End-to-end optimization of data-driven high-order Godunov-type finite-volume schemes for compressible homogeneous isotropic turbulence</em>”. Journal of Computational Physics 522(2025).</li><li>Nicola Clinco, Niccolò Tonicello and Gianluigi Rozza. “<em>A data-driven study on implicit LES using a spectral difference method</em>”. Journal of Computational Physics 540 (2025).</li><li>Nicola Clinco, Niccolò Tonicello, Paola Cinnella and Gianluigi Rozza. “<em>End-to-end optimization of subgrid scale models for discontinuous spectral element schemes based on the discrete adjoint method</em>”. arXiv:2606.01150 (2026).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43ee738c6c4e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/embedding-machine-learning-with-computational-fluid-dynamics-simulations-a-possible-marriage-43ee738c6c4e">Embedding Machine Learning with Computational Fluid Dynamics simulations: a possible marriage?</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Controlling Fluid Systems with Reinforcement Learning]]></title>
            <link>https://medium.com/sissa-mathlab/controlling-fluid-systems-with-reinforcement-learning-ffe7f68ff3c1?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/ffe7f68ff3c1</guid>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[fluid-dynamics]]></category>
            <category><![CDATA[optimal-control]]></category>
            <dc:creator><![CDATA[Elisa Savio]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 00:14:57 GMT</pubDate>
            <atom:updated>2026-07-12T00:14:55.409Z</atom:updated>
            <content:encoded><![CDATA[<h4>From Advection-Diffusion to Navier-Stokes</h4><p><a href="https://medium.com/u/deb5eb007cf3">Guglielmo Padula</a>, <a href="https://medium.com/u/6589d4a90499">Elisa Savio</a></p><h3>Introduction</h3><p>Reinforcement Learning (RL) is one of the three fields composing Machine Learning (ML). Unlike Supervised and Unsupervised Learning, it does not rely on the availability of data, but the <em>agent</em> is able to adjust its behavior, called <em>policy</em>, by self-interacting with an <em>environment</em>. The goal of RL is to adapt the policy in such a way to control the environment in a desired behavior. The environment returns a quantity, called <em>reward</em>, that indicates how well the policy is performing. For example, the agent may be a dog that is learning to pick up the stick thrown by its owner (the environment) and bring it back to him. Whenever he does it successfully, he gets food.</p><p>We present an application of Deep Reinforcement Learning (DRL) — specifically the Deep Deterministic Policy Gradient (DDPG) algorithm — to the problem of volumetric control of Partial Differential Equations (PDEs) arising in fluid mechanics. After a self-contained introduction to Markov Decision Processes (MDPs) and to the DDPG method, we describe two test cases: an advection-diffusion equation, for which an optimal Linear Quadratic Regulator (LQR) solution is available as a baseline, and the incompressible Navier-Stokes equations posed in a two dimensional cavity. In both cases, DDPG successfully learns a feedback control policy directly from environment interactions, without requiring any knowledge of the system matrices.</p><h3>Markov Decision Processes</h3><p>A Markov Decision Process [1] is a tuple (<em>S, A, P, r, γ</em>), where:</p><ul><li><em>S</em> is the state space,</li><li><em>A</em> is the action space,</li><li><em>P(s’ | s, a)</em> is the transition probability,</li><li><em>r: S</em> × <em>A → </em>ℝ is the reward function,</li><li><em>γ</em> ∈ [0,1) is the discount factor. The value <em>γ</em> = 1 is admissible only if the time horizon is finite.</li></ul><p>At each discrete time step <em>t </em>∈ {0, 1, … <em>T</em>}, the agent observes the current state <em>s_t</em> of the environment, selects action <em>a_t</em> according to policy π, receives reward <em>r_t = r(s_t, a_t)</em>, and transitions to state <em>s_{t+1} ~ P( · | s_t, a_t)</em>. The objective is to find a policy that maximizes the expected discounted cumulative reward:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CwI4HY7ylK4XzUY0fQItSw.png" /></figure><p>The Markov property — the fact that the future is conditionally independent of the past given the present state—as expressed in the following equation</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x7uunUKmeZ-FWikKkA3rVQ.png" /></figure><p>is what makes this framework tractable and distinguishes it from general history-dependent sequential decision problems.</p><p>Going back to the example of the dog and its owner:</p><ul><li>the state space consists of the coordinates of the dog, of the stick, and a boolean that indicates if the dog holds the stick or not;</li><li>the dog can move in four directions (up, down, left, right) or act on the stick (pickup, drop);</li><li>the dog is rewarded when it reaches the stick or when it brings back the stick to the owner.</li></ul><h3>Reinforcement Learning</h3><p>Reinforcement Learning [1] is the computational framework for finding good policies in an MDP when the transition probability <em>P</em> and/or the reward function <em>r</em> are unknown and must be learned through interaction with the environment.</p><p>Central to most modern RL algorithms is the notion of value functions. The action-value function, also called Q-function, of policy π is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qkjDBdTDNpwMHHVicl8myA.png" /></figure><p>and satisfies the so-called Bellman equation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nHFIWYFHFV6SlELqahWxUQ.png" /></figure><p>The goal is to find the optimal Q-function</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nzxDgg_zWADczvE07s2I_A.png" /></figure><p>from which the optimal deterministic policy follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nUyrQoW8rTD1QcsCGJzr4g.png" /></figure><p>Returning to our example, in Figure 1 the dog is still learning how to behave, since the optimal Q-function has not been computed yet. In Figure 2, after 300 episodes, the dog is able to reach the goal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*kUFLKmHI2mdGvA2D7cNvZw.gif" /><figcaption>Figure 1: before the optimal Q-function has been found, the dog is still learning how to behave.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*XqFLL__KZ3zIiAu5b0gQUQ.gif" /><figcaption>Figure 2: the dog follows a good policy, computed by the current Q-function, which is almost optimal.</figcaption></figure><p>When the state and action spaces are finite and have small dimensions, the Q-function can be stored and updated as a matrix: this is the case of tabular RL. However, in most of the real world applications, the state and action spaces have large dimensions, even infinite: therefore, the Q-function cannot be represented using a matrix, and is usually approximated using Deep Neural Networks (DNNs). This sub-field is called Deep Reinforcement Learning.</p><h3>Deep Deterministic Policy Gradient</h3><p>The Deep Deterministic Policy Gradient [2] algorithm combines the actor-critic architecture with deep neural network function approximators. In particular, DDPG maintains four networks:</p><ul><li><strong>Actor</strong> <em>μ_θ(s)</em>: a deterministic policy mapping states to actions.</li><li><strong>Critic </strong><em>Q_</em>φ<em>(s, a)</em>: an estimate of the action-value function.</li><li><strong>Target actor</strong> <em>μ_{θ’}</em> and <strong>target critic</strong> <em>Q_{</em>φ<em>’}</em>: slowly update copies used to stabilize training.</li></ul><p>Experience tuples <em>(s_t, a_t, r_t, s_{t+1})</em> are stored in a replay buffer <em>D</em> and sampled in random mini-batches. The critic is updated by minimising the mean-squared Bellman error:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SUEC05uxckDa_ShSknkQug.png" /></figure><p>while the actor is updated by ascending the sampled policy gradient:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TliqD4XEeKqykT1wuMhSEw.png" /></figure><p>Target networks are updated via a soft (Polyak) average:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H47TvtO4XdmB3omj9Up-zw.png" /></figure><p>Exploration is encouraged by adding Gaussian noise to the actor output during training.</p><h3>Test Case 1: Advection-Diffusion Equation</h3><h4>Problem Setting</h4><p>As a first, analytically tractable benchmark, we consider the advection-diffusion equation on the unit square Ω = [0,1]² with homogeneous Dirichlet boundary conditions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nc5_VqsSQs0-vTxu0bdJiw.png" /></figure><p>where:</p><ul><li><em>β </em>= (1,1)ᵀ is the convective velocity,</li><li><em>μ</em> = 1 is the diffusion coefficient,</li><li>φ_i(<strong>x</strong>) = exp(−‖<strong>x</strong> − <strong>x</strong>_i^c‖²) are Gaussian actuator functions centered at a grid of control points,</li><li><em>a_i(t)</em> are the control amplitudes.</li></ul><p>The initial condition is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OgNdA_WFpt74fAjiLME-Ug.png" /></figure><p>Applying classical stabilised finite element methods results in the following spatially discretized system, obtained via scikit-fem [5]:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ste0fRj2_R8KwTppS5Z6Eg.png" /></figure><p>where <em>M</em> is the mass matrix, <em>A</em> is the stifness-plus-advection matrix, and B is the control influence matrix. Implicit Euler time integration with step Δ<em>t</em> yields:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SROkMUtkdO_uieotzOIMag.png" /></figure><h4>Classical baseline: Linear Quadratic Regulator</h4><p>Since the discrete system is linear, an optimal controller can be computed analytically as the solution of a finite-horizion Linear Quadratic Regulator problem. The optimal feedback gain is obtained by solving the Riccati recursion backwards in time [3]:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1HdIF7p0Ktnne9b45maXjA.png" /></figure><p>with terminal condition <em>P_T</em> = Q = I and cost weights R. The controlled state then evolves as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LBP72nsYKNDgIkGFFJ0XMw.png" /></figure><p>where</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c6SD7Bphz5tSqBYsJ0MwYA.png" /></figure><h4>RL formulation</h4><p>The MDP for the advection-diffusion problem is defined by:</p><ul><li>State: <em>s_t</em> = <strong>u</strong><em>^t</em>.</li><li>Action: <em>a_t</em> ∈ [−l, l]^{N_0} (actuator amplitudes).</li><li>Reward: <em>r_t</em> = −(‖<strong>u</strong>^t‖² + (1/10)‖<strong>a</strong>^t‖²), matching the LQR cost.</li><li>Horizon: T = 50 steps.</li><li><em>γ</em> = 1 (undiscounted).</li></ul><p>DDPG is trained with Stable Baselines3 [4] for 10,000 environment time steps with learning rate equal to 5×10⁴.</p><h4>Results</h4><p>Figure 3 shows the evolution of ‖<strong>u</strong>^t‖ for the three following trajectories: uncontrolled, LQR-controlled, and DDPG-controlled. The DDPG policy achieves a norm reduction very close to that of the LQR solution, performing only marginally worse than the analytical optimum despite having no access to the system matrices. Both controlled trajectories vastly outperform the uncontrolled baseline, confirming that the Gaussian actuators provide sufficient controllability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*ltasC-hpZe9jKO9EmHbntg.png" /><figcaption>Figure 3</figcaption></figure><h3>Test Case 2: Incompressible Navier-Stokes Equations</h3><h4>Problem Setting</h4><p>The second, more challenging test case is the incompressible Navier-Stokes equations on Ω = [0,1]² with no-slip boundary conditions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KblzWTBCF2xKC02ZM30KIQ.png" /></figure><p>where <strong>u</strong> = (u₁, u₂)ᵀ is the velocity field, <em>p</em> is the pressure, and <em>ν</em> is the kinematic viscosity. The initial condition is <strong>u</strong>₀ = x(1−x)y(1−y)(1,1)ᵀ. There is no analytical optimal control available for this non-linear system.</p><h4>Finite element discretisation and POD reduction</h4><p>The velocity field is discretised with the Taylor–Hood-like pair (biquadratic Q₂ for velocity, bilinear Q₁ for pressure) on a 4-times-refined quadrilateral mesh, yielding n_u velocity and n_p pressure degrees of freedom. The non-linear convective term is assembled as a trilinear form and linearised around the current state. At each time step the coupled saddle-point system is solved:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K08aW_vK0bDqZ1zbCTDU5A.png" /></figure><p>with Δ<em>t</em> as time step. To keep the RL state space tractable, a Proper Orthogonal Decomposition (POD) reduction is applied: N_snap = 200 snapshots from an uncontrolled linearised rollout are collected, and the r = 30 leading left singular vectors Φ ∈ ℝ^{n_u × r} are retained as the POD basis. The propagation and control matrices are projected accordingly:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GIwweZ37o6jpHJ9_w89HNw.png" /></figure><h4>RL formulation</h4><p>The MDP for the Navier-Stokes problem is defined by:</p><ul><li>State: <em>s_t</em> = Φᵀu^t ∈ ℝ<em>^r</em> (POD coefficients).</li><li>Action: <em>a_t</em> ∈ [−1, 1]^N_c (actuator weights).</li><li>Reward: <em>r_t</em> = −(‖u_r^t − <strong>u</strong>_{r,0}‖² + α‖<strong>a</strong>^t‖²) with α = 10⁻³. The goal is to track the initial state.</li><li>Horizon: T = 50 steps.</li><li><em>γ</em> = 0.99.</li></ul><p>DDPG is trained for 100,000 environment steps using two hidden layers of 256 neurons and exploration noise σ = 0.1. The full-space error is reconstructed as ‖Φs_t − u₀‖ for evaluation.</p><h4>Results</h4><p>Figure 4 compares the full-space norm error ‖<strong>u</strong>(t) − <strong>u</strong>₀‖ between the uncontrolled and DDPG-controlled trajectories. The DDPG controller reliably lowers the norm error with respect to the uncontrolled flow, demonstrating that a policy learned entirely from reward signals in a low dimensional POD subspace can successfully regulate a high-dimensional nonlinear fluid system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*4mCpcmEIqBa8Hm3fZBWZFg.png" /><figcaption>Figure 4</figcaption></figure><h3>Conclusion</h3><p>We have shown that DDPG, a model-free deep RL algorithm designed for continuous action spaces, can successfully control PDE-governed dynamical systems without any access to the underlying system operators. In the linear advection–diffusion test case, DDPG nearly matches the analytically optimal LQR solution. In the more challenging nonlinear Navier–Stokes test case, combining DDPG with a POD-based state compression yields a controller that measurably attenuates the flow departure from the target state.</p><p>These results suggest that model-free RL is a viable complement to classical model-based control in scientific computing, particularly when the system equations are expensive to linearise or when the optimal control problem lacks a closed-form solution.</p><p>All source code is available at this <a href="https://github.com/guglielmopadula/RLMedium">link</a>.</p><h3>References</h3><p>[1] Sutton, R. S., and Barto, A. G., Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.</p><p>[2] Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D., “Continuous Control with Deep Reinforcement Learning,” in Proc. International Conference on Learning Representations (ICLR), 2016. arXiv:1509.02971.</p><p>[3] Kalman, R. E., “Contributions to the Theory of Optimal Control,” Boletín de la Sociedad Matemática Mexicana, vol. 5, pp. 102–119, 1960.</p><p>[4] Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., and Dormann, N., “Stable-Baselines3: Reliable Reinforcement Learning Implementations,” Journal of Machine Learning Research, vol. 22, no. 268, pp. 1–8, 2021.</p><p>[5] Gustafsson, T., and McBain, G. D., “scikit-fem: A Python Package for Finite Element Assembly,” Journal of Open Source Software, vol. 5, no. 52, p. 2369, 2020.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ffe7f68ff3c1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/controlling-fluid-systems-with-reinforcement-learning-ffe7f68ff3c1">Controlling Fluid Systems with Reinforcement Learning</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[dealii-X — Digital Twins of Human
Organs]]></title>
            <link>https://medium.com/sissa-mathlab/dealii-x-digital-twins-of-human-organs-de9be7f6b9b3?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/de9be7f6b9b3</guid>
            <category><![CDATA[digital-twin]]></category>
            <category><![CDATA[numerical-analysis]]></category>
            <category><![CDATA[high-performace-computing]]></category>
            <category><![CDATA[human-organ]]></category>
            <category><![CDATA[finite-element-method]]></category>
            <dc:creator><![CDATA[Pasquale Claudio Africa]]></dc:creator>
            <pubDate>Mon, 25 May 2026 01:37:26 GMT</pubDate>
            <atom:updated>2026-05-25T01:37:25.351Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Pasquale Claudio Africa, Andrea Cangiani, Marco Feder</em></p><p>Accurately modeling human organs is currently at the very limit of today’s computational capabilities, even when employing highly efficient and sophisticated numerical methods. Their biological functions involve highly complex geometries and multiscale phenomena (both in time and space), with fast signals between cells resulting in long-term consequences, such as the development of an illness.</p><p>Efficient and high-fidelity computational modeling of human organs is precisely the goal of the project <strong>dealii-X — Exascale Framework for Digital Twins of the Human Body</strong> (<a href="https://www.dealii-x.eu/">https://www.dealii-x.eu/</a>). The project is developing a high-performance, scalable computational platform, based on the deal.II library [2], that performs accurate digital twins of the brain, cardiovascular and respiratory systems, and the liver.</p><p>The dealii-X SISSA team is developing novel numerical techniques, such as reduced-order modeling, mesh adaptivity, and polytopic discretizations to mitigate the intrinsic complexity of biological processes. These techniques are tested within a comprehensive computational model of the human heart.</p><p>Mathematical modeling begins with reconstructing the organ tissues from medical imaging, such as computerized tomography (CT). An example is shown in the figure below, which depicts a real-life brain model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uKPB5KNjaQbNUFb4abYuyg.jpeg" /><figcaption>Tetrahedral mesh of a human brain geometry acquired from computerized tomography from [7].</figcaption></figure><p>The brain is represented in the computer as tens of millions of “pixels” (tetrahedra) that represent the geometry of the tissue as acquired from the CT scan. Performing simulations on such realistic geometries involves solving systems of equations with up to billions of variables multiple times. <em>Upscaling</em> techniques can be employed to reduce this prohibitive complexity. One such approach is that of agglomerating neighboring tetrahedral elements into larger polyhedral elements, thus reducing their number without compromising the resolution of the geometry (see the figure below). In turn, this considerable reduction in terms of the number of elements results in much smaller linear systems of equations to solve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OcfgYXXmCFkcTy5LRs_RZw.jpeg" /><figcaption>Portion of the brain geometry from [7]. A coarse polyhedral mesh is extracted by agglomerating together neighbouring tetrahedral elements.</figcaption></figure><h3>Polytopic Discretizations and Multilevel Methods</h3><p>Classical Finite Element Methods (FEM) matured in the mid-20th century with simplicial (triangular and tetrahedral) and tensor product elements (quadrilateral and hexahedral); its roots are usually traced to Hrennikoff (1941) and Courant (1943), and the success of the method across engineering and applied mathematics is well established.</p><p>From the early 2000s, a family of <em>polytopic</em> discretizations emerged, allowing meshes made of general polygons in 2D and polyhedra in 3D — with further extensions to curved elements also allowed. Such flexibility permits, for instance, the meshing of complex geometries and handling mesh distortion within dynamic problems and hanging nodes within adaptive mesh refinement.</p><p>Important strands include the conforming polygonal FEM via generalized barycentric/natural-neighbor bases [8], mimetic finite difference (MFD) in polytopic meshes [3], preserving symmetry/stability for diffusion and beyond, the Virtual Element Method (VEM) [4], which replaces explicit shape functions by polynomial projections and stabilizations, and the <em>hp</em>-version discontinuous Galerkin methods [6], which naturally extend to extremely general mesh discretizations. Other approaches include the general class of <em>unfitted</em> methods, designed to flexibly and accurately handle complicated, evolving, and moving geometries by decoupling the domain description from the underlying discretization [5].</p><p>The development of polytopic discretizations requires a careful mathematical analysis as well as efficient implementations. The potential benefit is the ability to perform numerical computations of mathematical models involving highly complex multi-scale phenomena at an unprecedented level of resolution.</p><p>A rigorous mathematical analysis has shown that discontinuous Galerkin methods allow the use of mesh shapes of unprecedented generality [6]. In particular, upscaling of complicated geometries becomes possible via mesh agglomeration as described above. This can be used either to perform coarser (cheaper) computations or within multigrid solvers, a preconditioning technique that accelerates the solution of the underlying problem by exploiting multiple levels of meshes.</p><p>In this realm, we introduced a novel multilevel agglomeration and preconditioning strategy designed to solve the monodomain equation in cardiac electrophysiology using high-order discontinuous Galerkin (DG) methods. At the core of this approach is the R*-tree spatial data structure, which is used to generate coarse grid hierarchies from existing fine meshes. By traversing tree levels, the method creates agglomerated polytopic elements, enabling efficient and robust multilevel computations [7].</p><p>Mathematically, the method leverages inherited bilinear forms, constructing coarse operators via Galerkin projection, which eliminates the costly step of assembling operators on complex polytopic elements. The monodomain equation</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/302/1*6rFRYMofufJPc1437MEDAQ.png" /></figure><p>is discretized using Symmetric Interior Penalty DG with semi-implicit time stepping, treating diffusion implicitly and reaction terms explicitly.</p><p>The C++ implementation builds on the <a href="https://github.com/fdrmrc/Polydeal">Polydeal</a> library, which extends the deal.II library capabilities to handle general polytopal shapes. Our agglomeration algorithm scales efficiently from 128 to 1024 processors, even on realistic three-dimensional cardiac geometries, while maintaining load balance. Memory consumption is minimized, when possible, through matrix-free implementations of prolongation and restriction operators.</p><p>Numerical experiments show clear performance advantages over traditional algebraic multigrid preconditioning. For polynomial degrees <em>p=1</em> to <em>p=5</em>, our method consistently reduces iteration counts, with gains increasing for higher-order polynomials (see figures below). Its performance is largely independent of the number of multigrid levels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6Wf0Rm1pmnKot8iKcGYpGw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m45gwurFYrxyPWkUUUlbfA.jpeg" /><figcaption>Ventricle electrophysiology model from [1]. The multigrid solver based on mesh agglomeration requires roughly half the iteration needed by state-of-the-art algebric multigrid solver, thus reducing computational time.</figcaption></figure><p>The coarsening algorithm preserves geometric quality and the spatial locality of agglomerated elements, resulting in well-conditioned coarse operators. The preconditioner integrates a V-cycle multigrid with Chebyshev smoothers, where eigenvalue estimates are efficiently computed using Lanczos iterations.</p><h3>References</h3><p>[1] P. C. Africa and M. Feder, <em>A polytopic agglomeration-based multigrid strategy for the discontinuous Galerkin discretization of cardiac electrophysiology</em>, In preparation.<br>[2] D. Arndt, W. Bangerth, D. Davydov, T. Heister, L. Heltai, M. Kronbichler, M. Maier, J.-P. Pelteret, B. Turcksin, and D. Wells, <em>The deal.II finite element library: Design, features, and insights</em>, <strong>Computers &amp; Mathematics with Applications</strong>, 81 (2021), pp. 407–422.<br>[3] L. C. Beirao da Veiga, K. Lipnikov, and G. Manzini, <em>The mimetic finite difference method for elliptic problems</em>, vol. 11 of <strong>MS&amp;A. Modeling, Simulation and Applications</strong>, Springer, Cham, 2014.<br>[4] L. Beirao da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L. Marini, and A. Russo, <em>Basic principles of virtual element methods</em>, <strong>Math. Models Methods Appl. Sci.</strong>, 23 (2013), pp. 199–214.<br>[5] D. Boffi, A. Cangiani, M. Feder, L. Gastaldi, and L. Heltai, <em>A comparison of non-matching techniques for the finite element approximation of interface problems</em>, <strong>Comput. Math. Appl.</strong>, 151 (2023), pp. 101–115.<br>[6] A. Cangiani, Z. Dong, E. H. Georgoulis, and P. Houston, <em>hp-version discontinuous Galerkin methods on polygonal and polyhedral meshes</em>, <strong>SpringerBriefs in Mathematics</strong>, Springer, Cham, 2017.<br>[7] M. Feder, A. Cangiani, and L. Heltai, <em>R3mg: R-tree based agglomeration of polytopal grids with applications to multilevel methods</em>, <strong>Journal of Computational Physics</strong>, 526 (2025), p. 113773.<br>[8] N. Sukumar and A. Tabarraei, <em>Conforming polygonal finite elements</em>, <strong>Internat. J. Numer. Methods Engrg.</strong>, 61 (2004), pp. 2045–2066.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de9be7f6b9b3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/dealii-x-digital-twins-of-human-organs-de9be7f6b9b3">dealii-X — Digital Twins of Human
Organs</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Introduction to Graph Neural Networks]]></title>
            <link>https://medium.com/sissa-mathlab/an-introduction-to-graph-neural-networks-7ed3ec0dd344?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/7ed3ec0dd344</guid>
            <category><![CDATA[graph-neural-networks]]></category>
            <category><![CDATA[geometric-deep-learning]]></category>
            <dc:creator><![CDATA[Giuseppe Alessio D'Inverno]]></dc:creator>
            <pubDate>Mon, 25 May 2026 01:37:06 GMT</pubDate>
            <atom:updated>2026-05-25T01:37:06.420Z</atom:updated>
            <content:encoded><![CDATA[<p>From social networks to molecular structures, many aspects of our world are best understood not as isolated points, but as intricate webs of interconnected entities. This is the realm of <strong>graphs</strong>, a powerful mathematical structure used to represent relationships.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/420/1*2-OrvkPo8echkTqPuBdPrA.png" /><figcaption>An example of graph [20].</figcaption></figure><p>Just to mention few examples:</p><ul><li><strong>Social media platforms</strong> like Facebook and X are essentially massive graphs where users are nodes and friendships or follows are edges.</li><li>In <strong>biology</strong>, protein-protein interaction networks, gene regulatory networks, and even the human brain’s neural connections are all graphs.</li><li><strong>Chemistry</strong> represents molecules as graphs, with atoms as nodes and chemical bonds as edges.</li><li><strong>Logistics and transportation</strong> rely on graphs to model road networks, public transit systems, and supply chains.</li><li>Even in <strong>e-commerce</strong>, recommendation systems can be viewed as graphs connecting users, products, and their interactions.</li></ul><p>The list goes on, highlighting the immense prevalence of structured data in the form of graphs across countless scientific and industrial applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4J5AJ_wvlPKNjKvV3eFJsg.jpeg" /><figcaption>A road network visualizable as graph [19].</figcaption></figure><p><strong>The Challenge of Traditional Graph Algorithms</strong></p><p>For decades, computer scientists have developed sophisticated <strong>combinatorial algorithms</strong> to tackle problems on graphs. Algorithms like Dijkstra’s for shortest paths, PageRank for node importance, or various clustering algorithms have been instrumental in solving complex challenges.</p><p>However, these traditional methods often hit a wall when faced with real-world scenarios:</p><ul><li><strong>Scalability issues:</strong> Many graph problems are NP-hard, meaning the time required to find an optimal solution grows exponentially with the size of the graph. Real-world graphs can have billions of nodes and edges, making exact solutions computationally intractable.</li><li><strong>Feature engineering:</strong> Traditional algorithms often require handcrafted features to represent nodes and edges, which can be time-consuming, domain-specific, and may not capture the underlying complexities of the data.</li><li><strong>Generalization:</strong> Algorithms designed for specific graph structures may not generalize well to variations or entirely different types of graphs.</li></ul><p>This inefficiency and lack of adaptability have spurred the search for more powerful and flexible approaches.</p><p><strong>Deep Learning to the Rescue: Enter Graph Neural Networks</strong></p><p>The remarkable success of <strong>deep learning</strong> in fields like computer vision and natural language processing, where it excels at learning representations from unstructured data (images, text), naturally led researchers to wonder: <em>Can we apply deep learning to structured data like graphs?</em></p><p>This question marked the birth of <strong>Graph Neural Networks (GNNs)</strong>. Instead of relying on predefined features or fixed algorithms, GNNs aim to learn meaningful representations (embeddings) of nodes and edges directly from the graph structure and associated features, enabling them to perform various tasks like classification, prediction, and generation.</p><p><strong>What is a Graph Neural Network? A Brief History</strong></p><p>At its core, a <strong>Graph Neural Network</strong> is a class of deep learning methods designed to operate directly on graph-structured data. Unlike traditional neural networks that operate on flat vectors or grids, GNNs are specifically designed to handle the irregular topology of graphs.</p><p>The concept of GNNs can be traced back to the early 2000s, with pioneering work by Franco Scarselli and his colleagues [1]. Their initial formulation laid the groundwork for learning node representations by iteratively propagating information across the graph.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5C9XDMdnBrQlQV6gTYqZUQ.png" /><figcaption>The original GNN model introduced in [1].</figcaption></figure><p>However, GNNs truly gained momentum in the mid-2010s with the advent of more scalable and effective architectures, largely inspired by the success of convolutional neural networks (CNNs) in image processing [2]. This led to the development of various GNN variants, each with its unique approach to aggregating and transforming information [3].</p><p><strong>The Message Passing Framework: The Heart of GNNs</strong></p><p>The majority of modern GNNs operate under a unified paradigm known as <strong>message passing </strong>[4]. This framework provides an intuitive and powerful way for nodes in a graph to learn about their neighborhood and the overall graph structure.</p><p>Imagine each node in a graph as an individual. In the message passing framework, these individuals communicate with their immediate neighbors in a series of rounds. In each round (or layer of the GNN), a node updates its representation (its “state” or “embedding”) by performing two main steps:</p><ol><li><strong>Message Aggregation:</strong> Each node collects “messages” from its direct neighbors. These messages are typically transformations of the neighbors’ current representations. The aggregation function (e.g., sum, mean, max) combines these incoming messages into a single aggregated message.</li><li><strong>Message Update:</strong> The node then combines this aggregated message with its own current representation (and potentially its own initial features) to compute its new, updated representation. This update often involves a neural network layer, like an MLP or a recurrent unit.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/1*rPGq0o2sAY2lEXNR3msvtg.jpeg" /><figcaption>The message passing iteration, as introduced in [4].</figcaption></figure><p>This process is repeated for several layers. With each successive layer, a node’s representation incorporates information from increasingly distant neighbors, effectively expanding its “receptive field” across the graph. After several message passing iterations, the final representations of the nodes are rich embeddings that encode both their local neighborhood information and their position within the broader graph structure.</p><p>The node embeddings learned with GNNs can then be used for various downstream tasks:</p><ul><li><strong>Node classification:</strong> Predicting the type or category of a node.</li><li><strong>Link prediction:</strong> Predicting the existence of a missing link between two nodes.</li><li><strong>Graph classification:</strong> Classifying an entire graph based on its structure and features.</li><li><strong>Node clustering:</strong> Grouping similar nodes together.</li></ul><p><strong>The Connection to the Weisfeiler-Lehman (1-WL) Test</strong></p><p>A crucial theoretical insight into the expressive power of many message-passing GNNs comes from their connection to the <strong>Weisfeiler-Lehman (1-WL) test</strong> for graph isomorphism [5]. The 1-WL test, also known as the color refinement algorithm, is a classic heuristic for determining if two graphs are isomorphic (structurally identical).</p><p>Here’s how the 1-WL test works:</p><ol><li><strong>Initialization:</strong> All nodes are initially assigned a color (or label) based on their features (or a default color if no features are present).</li><li><strong>Iterative Refinement:</strong> In each iteration, every node computes a new color based on its <em>current color</em> and a <em>multiset of the colors of its neighbors</em>. This is typically done using an injective (collision-free) hash function.</li><li><strong>Convergence:</strong> The process repeats until the node colors no longer change, or for a fixed number of iterations. If two graphs can be distinguished by the 1-WL test (i.e., they produce different sets of colors or color histograms),then they are non-isomorphic.</li></ol><p><strong>The remarkable connection is that many standard message-passing GNNs are proven to be at most as powerful as the 1-WL test in distinguishing non-isomorphic graphs [6,7].</strong> This means that if the 1-WL test cannot distinguish between two graphs, a typical message-passing GNN will also likely produce similar embeddings for them, regardless of the complexity of the neural network layers used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/614/1*BC5BcfU2wBskh4ZsPODh2w.png" /><figcaption>The iterative update of the 1-WL test. Note that it coincides with the notion of <strong>unfolding equivalence </strong>[13].</figcaption></figure><p>This insight provides a theoretical upper bound on the expressive power of these GNNs. While the 1-WL test is very efficient and can distinguish a vast majority of real-world non-isomorphic graphs, it does have limitations (e.g., it fails for certain highly symmetric graphs). This understanding has spurred research into more expressive GNN architectures that aim to go “beyond 1-WL” by incorporating higher-order information or other mechanisms to overcome these limitations [8,9,10,11].</p><p>From a theoretical point of view, GNNs have been characterized also in terms of approximation universality[12,13,14] and generalization capabilities[15,16,17,18].</p><p>We can see how to practically implement a GNN using the <strong>Pytorch Geometric </strong>framework [21] as in the snippet below:</p><pre>import torch<br>import torch.nn.functional as F<br>from torch_geometric.nn import GraphConv<br>from torch_geometric.data import Data<br><br># 1. Prepare your Graph Data<br># Let&#39;s create a simple graph with 4 nodes and some features and edges.<br><br># Node features (e.g., degree, color, etc.)<br># Here, each node has a 2-dimensional feature vector.<br>x = torch.tensor([[1.0, 2.0],<br>                  [3.0, 4.0],<br>                  [5.0, 6.0],<br>                  [7.0, 8.0]], dtype=torch.float)<br><br># Edge list (connections between nodes)<br># This represents an undirected graph. For directed, specify (src, dest)<br># [[0, 1], [1, 0], [1, 2], [2, 1], [2, 3], [3, 2]]<br>edge_index = torch.tensor([[0, 1, 1, 2, 2, 3],<br>                           [1, 0, 2, 1, 3, 2]], dtype=torch.long)<br><br># Node labels (what we want to predict for each node)<br># For a node classification task, each node has a label.<br>y = torch.tensor([0, 1, 0, 1], dtype=torch.long).unsqueeze(-1)<br><br># Create a PyTorch Geometric Data object<br>data = Data(x=x, edge_index=edge_index, y=y)<br><br>print(&quot;--- Graph Data ---&quot;)<br>print(data)<br>print(f&quot;Number of nodes: {data.num_nodes}&quot;)<br>print(f&quot;Number of edges: {data.num_edges}&quot;)<br>print(f&quot;Node features shape: {data.x.shape}&quot;)<br>print(f&quot;Edge index shape: {data.edge_index.shape}&quot;)<br>print(f&quot;Node labels: {data.y}\n&quot;)<br><br># 2. Define your GNN Model using GraphConv<br>class GNN(torch.nn.Module):<br>    def __init__(self, in_channels, hidden_channels, out_channels):<br>        super().__init__()<br>        # GraphConv layer 1: Takes input features and transforms to hidden_channels<br>        self.conv1 = GraphConv(in_channels, hidden_channels)<br>        # GraphConv layer 2: Takes hidden_channels and transforms to out_channels<br>        # (typically the number of classes for node classification)<br>        self.conv2 = GraphConv(hidden_channels, out_channels)<br><br>    def forward(self, x, edge_index):<br>        # Apply the first GraphConv layer<br>        # x: Node features<br>        # edge_index: Graph connectivity<br>        x = self.conv1(x, edge_index)<br>        # Apply a non-linear activation function (ReLU is common)<br>        x = F.relu(x)<br>        # Apply the second GraphConv layer<br>        x = self.conv2(x, edge_index)<br>        # For node classification, typically apply log_softmax for probability distribution<br>        return torch.sigmoid(x)<br><br># 3. Initialize the Model, Optimizer, and Loss Function<br># input features: data.num_node_features (2 in our case)<br># hidden channels: arbitrary choice<br># output channels: data.num_classes (number of unique labels in y, 2 in our case: 0 and 1)<br>in_channels = data.num_node_features<br>hidden_channels = 4<br>out_channels = 1 # Number of unique classes<br><br>model = GNN(in_channels, hidden_channels, out_channels)<br>optimizer = torch.optim.Adam(model.parameters(), lr=0.001)<br>criterion = torch.nn.BCELoss() # For multi-class classification<br><br>print(&quot;--- Model Architecture ---&quot;)<br>print(model)<br>print(f&quot;Input channels: {in_channels}&quot;)<br>print(f&quot;Hidden channels: {hidden_channels}&quot;)<br>print(f&quot;Output channels (number of classes): {out_channels}\n&quot;)<br><br># 4. Training Loop<br>def train():<br>    model.train() # Set the model to training mode<br>    optimizer.zero_grad() # Clear gradients<br>    out = model(data.x, data.edge_index) # Forward pass<br>    loss = criterion(out, data.y.float()) # Compute loss<br>    loss.backward() # Backpropagate<br>    optimizer.step() # Update weights<br>    return loss.item()<br><br>print(&quot;--- Training ---&quot;)<br>for epoch in range(2000): # Train for 200 epochs<br>    loss = train()<br>    if (epoch + 1) % 20 == 0:<br>        print(f&#39;Epoch: {epoch+1:03d}, Loss: {loss:.4f}&#39;)<br><br># 5. Inference/Evaluation<br>def evaluate():<br>    model.eval() # Set the model to evaluation mode<br>    with torch.no_grad(): # Disable gradient computation<br>        out = model(data.x, data.edge_index)<br>        pred = out.int() # Get the predicted class<br>        correct = (pred == data.y).sum().item() # Count correct predictions<br>        accuracy = correct / data.num_nodes # Calculate accuracy<br>    return pred, accuracy<br><br>print(&quot;\n--- Evaluation ---&quot;)<br>predictions, accuracy = evaluate()<br>print(f&quot;True labels: {data.y.tolist()}&quot;)<br>print(f&quot;Predictions: {predictions.tolist()}&quot;)<br>print(f&quot;Accuracy: {accuracy:.4f}&quot;)</pre><p>In conclusion, by leveraging the message passing framework, and understanding its theoretical relationship to the 1-WL test, Graph Neural Networks offer a flexible and powerful approach to learning on structured data, opening up new possibilities for solving complex problems across diverse domains. As GNN research continues to evolve, we can expect even more innovative applications that harness the power of connections.</p><p><strong>References</strong></p><p>[1] Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., &amp; Monfardini, G. (2008). The graph neural network model. <em>IEEE transactions on neural networks</em>, <em>20</em>(1), 61–80.</p><p>[2] Kipf, T. N., &amp; Welling, M. (2017, February). Semi-Supervised Classification with Graph Convolutional Networks. In <em>International Conference on Learning Representations</em>.</p><p>[3] Corso, G., Stark, H., Jegelka, S., Jaakkola, T., &amp; Barzilay, R. (2024). Graph neural networks. <em>Nature Reviews Methods Primers</em>, <em>4</em>(1), 17.</p><p>[4] Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., &amp; Dahl, G. E. (2017, July). Neural message passing for quantum chemistry. In <em>International conference on machine learning</em> (pp. 1263–1272). PMLR.</p><p>[5] Weisfeiler, B., &amp; Leman, A. (1968). The reduction of a graph to canonical form and the algebra which appears therein. <em>nti, Series</em>, <em>2</em>(9), 12–16.</p><p>[6] Xu, K., Hu, W., Leskovec, J., &amp; Jegelka, S. How Powerful are Graph Neural Networks?. In <em>International Conference on Learning Representations</em>.</p><p>[7] Morris, C., Ritzert, M., Fey, M., Hamilton, W. L., Lenssen, J. E., Rattan, G., &amp; Grohe, M. (2019, July). Weisfeiler and leman go neural: Higher-order graph neural networks. In <em>Proceedings of the AAAI conference on artificial intelligence</em> (Vol. 33, №01, pp. 4602–4609).</p><p>[8] Chamberlain, B., Rowbottom, J., Gorinova, M. I., Bronstein, M., Webb, S., &amp; Rossi, E. (2021, July). Grand: Graph neural diffusion. In <em>International conference on machine learning</em> (pp. 1407–1418). PMLR.</p><p>[9] Di Giovanni, F., Rowbottom, J., Chamberlain, B. P., Markovich, T., &amp; Bronstein, M. M. (2022). Graph neural networks as gradient flows: understanding graph convolutions via energy. <em>arXiv preprint arXiv:2206.10991</em>.</p><p>[10] Topping, J., Di Giovanni, F., Chamberlain, B. P., Dong, X., &amp; Bronstein, M. M. Understanding over-squashing and bottlenecks on graphs via curvature. In <em>International Conference on Learning Representations</em>.</p><p>[11] Errica, F., Christiansen, H., Zaverkin, V., Maruyama, T., Niepert, M., &amp; Alesiani, F. (2023). Adaptive message passing: A general framework to mitigate oversmoothing, oversquashing, and underreaching. <em>arXiv preprint arXiv:2312.16560</em>.</p><p>[12] Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., &amp; Monfardini, G. (2008). Computational capabilities of graph neural networks. <em>IEEE Transactions on Neural Networks</em>, <em>20</em>(1), 81–102.</p><p>[13] D’Inverno, G. A., Bianchini, M., Sampoli, M. L., &amp; Scarselli, F. (2024). On the approximation capability of GNNs in node classification/regression tasks. <em>Soft Computing</em>, <em>28</em>(13), 8527–8547.</p><p>[14] Loukas, A. (2019). What graph neural networks cannot learn: depth vs width. <em>arXiv preprint arXiv:1907.03199</em>.</p><p>[15] Garg, V., Jegelka, S., &amp; Jaakkola, T. (2020, November). Generalization and representational limits of graph neural networks. In <em>International conference on machine learning</em> (pp. 3419–3430). PMLR.</p><p>[16] Scarselli, F., Tsoi, A. C., &amp; Hagenbuchner, M. (2018). The vapnik–chervonenkis dimension of graph and recursive neural networks. <em>Neural Networks</em>, <em>108</em>, 248–259.</p><p>[17] D’Inverno, G. A., Bianchini, M., &amp; Scarselli, F. (2025). VC dimension of Graph Neural Networks with Pfaffian activation functions. <em>Neural Networks</em>, <em>182</em>, 106924.</p><p>[18] Morris, C., Geerts, F., Tönshoff, J., &amp; Grohe, M. (2023, July). Wl meet vc. In <em>International conference on machine learning</em> (pp. 25275–25302). PMLR.</p><p>[19] Yan, L., Hu, W., &amp; Hu, S. (2019). SALA: a self-adaptive learning algorithm — towards efficient dynamic route guidance in urban trafficnetworks. <em>Neural Processing Letters</em>, <em>50</em>, 77–101.</p><p>[20] Chunaev, P. (2020). Community detection in node-attributed social networks: a survey. <em>Computer Science Review</em>, <em>37</em>, 100286.</p><p>[21] Fey, M., &amp; Lenssen, J. E. (2019). Fast graph representation learning with PyTorch Geometric. <em>arXiv preprint arXiv:1903.02428</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ed3ec0dd344" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/an-introduction-to-graph-neural-networks-7ed3ec0dd344">An Introduction to Graph Neural Networks</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data, Reduced Order Models and Artificial Intelligence for Digital Twins]]></title>
            <link>https://medium.com/sissa-mathlab/data-reduced-order-models-and-artificial-intelligence-for-digital-twins-b28d984140e4?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/b28d984140e4</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[digital-twin]]></category>
            <category><![CDATA[reduced-order-modeling]]></category>
            <dc:creator><![CDATA[Laura Meneghetti]]></dc:creator>
            <pubDate>Mon, 25 May 2026 01:35:03 GMT</pubDate>
            <atom:updated>2026-05-25T01:35:02.900Z</atom:updated>
            <content:encoded><![CDATA[<p>In recent years, an increasing number of disruptive innovations with significant economic and social impact are characterizing the digitalization processes. However, despite the benefits of these advancements, their speed and extent are constrained by the limitations of available simulation technologies in handling complex models. The core concept of the Digital Twin (DT) lies in creating a digital counterpart of a physical asset capable of replicating its behavior. The need for continuous and fast interaction between the physical and digital environments, supported by background simulations, has led to the adoption of Reduced Order Modeling (ROM) and Artificial Intelligence (AI) as key technologies. These approaches speed up model execution while maintaining a high level of accuracy with a reduced number of degrees of freedom. The synergy of these technologies with Big Data ensures more accurate and efficient decision-making, enhancing the performance and reliability of systems across various applications. By leveraging these techniques, DT can be applied to new and complex emerging fields, providing more precise insights into its real-world counterpart [15].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uVnqn2o4YtdCoy9K.jpg" /><figcaption><strong>Figure1</strong>: Digital Twin of an airplane</figcaption></figure><p><strong>An Integrated paradigm: Digital Twin</strong></p><p>A Digital Twin (DT) represents a virtual replica of an object or phenomena, which can be used to gain improvements in the physical part. This is obtained through an automatic flow of data between the physical object and the digital one. As represented in Figure 2, a DT is thus composed of three main parts:</p><ul><li>a Physical Object</li><li>a Virtual or Digital Object</li><li>Data</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*TB-dfb5rpdpzbpYY4L_bCA.png" /><figcaption><strong>Figure 2</strong>: Main components of a Digital Twin</figcaption></figure><p>The continuous exchange of data between the digital twin and the real counterpart makes thus DT a fundamental support tool during several phases of an object life-cycle:</p><ul><li>Design Phase (Virtual Prototype)</li><li>Product operating life (predictive maintenance)</li><li>Production cycle (quality improvement)</li></ul><p>It is therefore evident that a Digital Twin is based on a vast amount of data and complex simulations, requiring several techniques to handle <strong>Big data</strong> and to speed up model execution. These methods are represented by <strong>High Performance Computing (HPC)</strong>, <strong>Reduced Order Models (ROMs)</strong>, <strong>Artificial Intelligence (AI)</strong>, and <strong>Uncertainty Quantification (UQ)</strong>. As represented in Figure 3, DT can thus be described as a unitary paradigm integrating all these elements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fqqur6CK_GhIbrcW" /><figcaption><strong>Figure 3</strong>: An integrated paradigm: Digital Twin</figcaption></figure><p>Let us delve into the details to understand how each of these elements is integrated within the DT framework:</p><ol><li><strong>Big Data: </strong>Data are the basis for the construction of such models. Nowadays, IoT devices — e.g. sensors, actuators, gadgets, appliances, or machines — represent a great source of data due to their increasing presence in everyday objects, factories, etc. This results in a large amount of available data requiring advanced tools and techniques for storage, processing, and analysis to uncover valuable insights and drive decision-making. AI coupled with HPC allows this management of Big Data and a real-time analysis of the continuous flow of such data coming from various sources such as simulations, sensors, etc.</li><li><strong>High Performance Computing (HPC): </strong>This huge amount of data needs to be handled through suitable infrastructures able to capture the underlying structure and information contained inside them. These data must then be provided as input to mathematical models, characterized by a large number of parameters due to the complexity of the modeled problem. HPC is thus fundamental in storing and managing Big Data, and then running complex simulations.</li><li><strong>Reduced Order Models (ROMs)</strong>: Due to the complexity (high number of parameters) of models, simulations can be very slow and computationally expensive. To be applied in an industrial environment, they need to be real-time and with a low computational cost. ROMs represent thus an approach to handle this problem, providing models reduced in the number of parameters but with high accuracy. By detecting the most important directions and information in the parameter space, ROMs can reduce the model by projecting it onto a low-dimensional space determined by the aforementioned analysis. An approximated solution can then be constructed through an interpolation/regression process. Figure 4 provides a schematic representation of the reduction process of ROMs [14].</li><li><strong>Artificial Intelligence (AI)</strong>: AI techniques, such as neural networks, enhance simulations by learning complex patterns from data, improving accuracy, efficiency and generalizability. Therefore, integrating neural networks into ROMs offers a powerful approach to approximate complex systems efficiently and accurately.</li><li><strong>Uncertainty Quantification (UQ)</strong>: UQ involves evaluating and managing uncertainties from both data and the mathematical models used to interpret it. This process is crucial for enhancing the reliability and applicability of models across various fields.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*p3U9t9K81-dkUq25" /><figcaption><strong>Figure 4</strong>: Schematic representation of the reduction process for ROMs</figcaption></figure><p><strong>Digital Twin and Scientific Machine Learning</strong></p><p>Scientific Machine Learning (SciML) is a powerful tool for improving scientific computing that integrates physics-based and data-driven models for the numerical approximation of differential problems [16]. <strong>Physics-based models</strong> relies on deriving mathematical models from fundamental principles and solving them using numerical methods and computational algorithms. They involve experimentation, theory formulation, numerical simulations and approximation, and model validation. In contrast, <strong>data-driven models</strong> focus on identifying relationships between input and output data without assuming any underlying causality and inferring underlying physical principles. Recently, they have gained widespread attention, driven by the availability of <strong>big data</strong>, increasingly affordable computing power, and advancements in machine learning algorithms.</p><p>These two approaches can be combined into <strong>Scientific Machine Learning (SciML), </strong>a framework that leverages the strengths of both, integrating the foundational knowledge of traditional models (physical insights) with the efficiency and pattern-recognition strengths of machine learning. SciML enhances machine learning models with physical knowledge, improving accuracy and interpretability, while AI techniques can optimize and accelerate traditional computational models.</p><p>By integrating physics-based modeling and data-driven learning, SciML also plays a key role in developing <strong>digital twins</strong>, enabling real-time interaction between physical systems and their virtual counterparts. Figure 5 provides an overview of the described framework, highlighting the connections between the different topics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nl0c1YPZWbkfWFcBZ6jh3w.png" /><figcaption><strong>Figure 5</strong>: Overview of the framework</figcaption></figure><p>For a given problem (e.g., fluid dynamics or stochastic PDEs), a structured pipeline can be established to integrate Machine Learning into the solution of the corresponding PDEs, as summarized in Figure 6. The first step involves generating data relevant to the problem, which can be obtained through high-fidelity simulations or scattered domain data collection.</p><p>Once the problem and data are defined, a Machine Learning model can be constructed using approaches such as Neural Networks, POD with Interpolation, or Neural Operators. These techniques provide different strategies for developing ML-based models.</p><p>The final step focuses on optimizing the model through methods such as Supervised Learning, Physics-Informed loss functions, and gradient descent algorithms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*19cwk5XemJFDtede" /><figcaption><strong>Figure 6</strong>: Machine Learning pipeline</figcaption></figure><p>An example of integration of Deep Learning into PDEs is represented then by <strong>Physics-Informed Neural Networks (PINNs)</strong> [9,10, 11,12], an optimization technique that leverages Neural Networks to solve differential equations. In this framework, PDEs are solved by embedding physical laws and symmetries into the network. PINNs thus provide a balanced approach that integrates data and physical knowledge from ROM simulations, enabling the development of accurate and reliable ML models.</p><p><strong>Reduced Order Models enhanced by Deep Learning and Scientific Machine Learning</strong></p><p>As discussed in the previous sections, Artificial Intelligence is becoming an increasingly integrated tool across various fields. Specifically, it can be used to improve traditional data-driven ROM techniques in Computational Fluid Dynamics. Examples of such methods include:</p><ul><li><strong>Approximation in Reduced Order Model</strong>: Artificial Neural Networks (ANNs) are employed to extract information from data and approximate the functional relationship between input parameters and output solutions. Techniques such as POD-NN and AE-NN have been developed for this purpose [1, 2, 3, 4, 8]. By enhancing generalization and incorporating a multi-fidelity approach, Neural Networks can improve the accuracy of POD-based models.</li><li><strong>Auto-encoders for dimensionality reduction and manifold learning: </strong>Autoencoders are integrated into data-driven methods to capture a latent representation of the dynamical system. To incorporate physical knowledge of the model, a nonlinear manifold approach [6] can be combined with hyper-reduction, achieved through reduced over-collocation and teacher–student training of a reduced decoder [5].</li><li><strong>Reduction in wide parameter space by means of deep learning parameter domain decomposition: </strong>A Deep Learning framework can be designed to generalize a system’s evolution across initial conditions within an extended parameter space. In [7], a two-stage memory-aware ANN-based model order reduction approach was developed to address this challenge. The curse of dimensionality is initially mitigated by partitioning the parameter space and averaging ROM solutions. Subsequently, Long Short-Term Memory networks (LSTMs) are combined with Convolutional Networks (C-LSTM) to capture temporal correlations.</li><li><strong>Generative models to quantify model uncertainty</strong>: As outlined in the previous section, UQ is a crucial aspect to consider. In [13], an approach combining generative modeling with UQ is introduced, where generative models are used to learn probability distributions from the data and capture their intrinsic characteristics. This probabilistic approach enables thus a priori UQ quantification.</li></ul><p><strong>Applications of Digital Twin</strong></p><p>As discussed in previous sections, a digital twin is a virtual replica of a physical asset. This concept can be applied across various fields, encompassing both products and processes. For instance, a digital twin can represent a manufacturing asset, an entire production plant, an environment, or even complex workflows.</p><p>We developed three Digital Twins — specifically, three Live Demos — within the SMACT Odyssea Competence Center for Industry 4.0 in Northeast Italy. These serve as tangible examples to showcase the tool’s potential:</p><ol><li><strong>Smart Hybrid Energy Management System (</strong>developed with Wärtsilä): in this project we aim to construct a DT for hybrid vessel engines in order to <strong>identify deviations </strong>from the engine ideal behavior, already at test time.</li><li><strong>Heavy-Duty Machining Process Optimization (</strong>created with Danieli Automation): the goal of the project is to enhance machining efficiency by introducing a DT for fault detection and predictive maintenance.</li><li><strong>Aspiring Hood Digital Twin (</strong>currently under development with Electrolux): also in this project we aim to develop a DT to improve the product quality and reduce risk of malfunctions.</li></ol><p>Additionally, we highlight the work of <strong>iNEST Spoke 9</strong>, the Interconnected Innovation Ecosystem of Northeast Italy, which focuses on computational technologies for digital twins, including infrastructures, services, networks, communities, and applications in medicine. At the core of this innovative tool is the continuous exchange of information between the physical asset and its virtual replica. This real-time data flow is crucial for making informed decisions and taking timely actions when needed.</p><h3>References:</h3><p>[1] San, O., Maulik, R., &amp; Ahmed, M. (2019). <em>An artificial neural network framework for reduced order modeling of transient flows.</em> Communications in Nonlinear Science and Numerical Simulation, 77, 271–287. <a href="https://doi.org/10.1016/j.cnsns.2019.04.025">https://doi.org/10.1016/j.cnsns.2019.04.025</a></p><p>[2] Hesthaven, J. S., &amp; Ubbiali, S. (2018). <em>Non-intrusive reduced order modeling of nonlinear problems using neural networks</em>. Journal of Computational Physics, 363, 55–78. <a href="https://doi.org/10.1016/j.jcp.2018.02.037">https://doi.org/10.1016/j.jcp.2018.02.037</a></p><p>[3] Papapicco, D., Demo, N., Girfoglio, M., Stabile, G., &amp; Rozza, G. (2022). <em>The neural network shifted-proper orthogonal decomposition: a machine learning approach for non-linear reduction of hyperbolic equations.</em> Computer<em> </em>Methods in Applied Mechanics and Engineering, 392, 114687. <a href="https://doi.org/10.1016/j.cma.2022.114687">https://doi.org/10.1016/j.cma.2022.114687</a></p><p>[4] Shah, N. V., Girfoglio, M., Quintela, P., Rozza, G., Lengomin, A., Ballarin, F., &amp; Barral, P. (2022). <em>Finite element based Model Order Reduction for parametrized one-way coupled steady state linear thermo-mechanical problems</em>. Finite Elements in Analysis and Design, 212, 103837. <a href="https://doi.org/10.1016/j.finel.2022.103837">https://doi.org/10.1016/j.finel.2022.103837</a></p><p>[5] Romor, F., Stabile, G. &amp; Rozza, G.<em> </em>(2023)<em> Non-linear Manifold Reduced-Order Models with Convolutional Autoencoders and Reduced Over-Collocation Method</em>. Journal of Scientific Computing <strong>94</strong>, 74. <a href="https://doi.org/10.1007/s10915-023-02128-2">https://doi.org/10.1007/s10915-023-02128-2</a></p><p>[6] Lee, K., &amp; Carlberg, K. T. (2021). <em>Deep conservation: A latent-dynamics model for exact satisfaction of physical conservation laws</em>. In Proceedings of the AAAI Conference on Artificial Intelligence, 35, 277–285. <a href="https://doi.org/10.1609/aaai.v35i1.16102">https://doi.org/10.1609/aaai.v35i1.16102</a></p><p>[7] Gonnella, I. C., Hess, M. W., Stabile, G., &amp; Rozza, G. (2023). <em>A two-stage deep learning architecture for model reduction of parametric time-dependent problems</em>. Computers &amp; Mathematics with Applications, 149, 115–127. <a href="https://doi.org/10.1016/j.camwa.2023.08.026">https://doi.org/10.1016/j.camwa.2023.08.026</a></p><p>[8] Demo, N., Tezzele, M. &amp; Rozza, G. (2023). <em>A DeepONet multi-fidelity approach for residual learning in reduced order modeling</em>. Advanced Modeling and Simulation in Engineering Sciences 10, 12 . <a href="https://doi.org/10.1186/s40323-023-00249-9">https://doi.org/10.1186/s40323-023-00249-9</a></p><p>[9] Raissi, M., Perdikaris, P., and Karniadakis, G.E. (2019) <em>Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations</em>. Journal of Computational physics, 378, 686–707. <a href="https://doi.org/10.1016/j.jcp.2018.10.045">https://doi.org/10.1016/j.jcp.2018.10.045</a></p><p>[10] Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., &amp; Piccialli, F. (2022) <em>Scientific machine learning through physics–informed neural networks: Where we are and what’s next.</em> Journal of Scientific Computing 92, 88. <a href="https://doi.org/10.1007/s10915-022-01939-z">https://doi.org/10.1007/s10915-022-01939-z</a></p><p>[11] Demo, N., Strazzullo, M., &amp; Rozza, G. (2023). <em>An extended physics informed neural network for preliminary analysis of parametric optimal control problems.</em> Computers &amp; Mathematics with Applications, 143, 383–396. <a href="https://doi.org/10.1016/j.camwa.2023.05.004">https://doi.org/10.1016/j.camwa.2023.05.004</a></p><p>[12] Coscia, D., Ivagnes, A., Demo, N., &amp; Rozza, G. (2023). <em>Physics-Informed Neural networks for Advanced modeling</em>. Journal of Open Source Software, 8(87), 5352. <a href="https://doi.org/10.21105/joss.05352">https://doi.org/10.21105/joss.05352</a></p><p>[13] Coscia, D., Demo, N., &amp; Rozza, G. (2024). <em>Generative Adversarial Reduced Order Modelling</em>. Nature Scientific Report, <strong>14</strong>, 3826.<em> </em><a href="https://doi.org/10.1038/s41598-024-54067-z">https://doi.org/10.1038/s41598-024-54067-z</a></p><p>[14] Rozza, G., Stabile, G., &amp; Ballarin, F. (Eds.). (2022). <em>Advanced reduced order methods and applications in computational fluid dynamics</em>. Society for Industrial and Applied Mathematics. <a href="https://doi.org/10.1137/1.9781611977257">https://doi.org/10.1137/1.9781611977257</a></p><p>[15] Rozza, G., Meneghetti, L., &amp; Martini, A. (2024) <em>Data, Reduced Order Models and Artificial Intelligence for Digital Twin. </em>Atti della Conferenza GARR 2024-Navigare la complessità. <a href="https://doi.org/2024.10.26314/GARR-Conf24-proceedings-10">https://doi.org/2024.10.26314/GARR-Conf24-proceedings-10</a></p><p>[16] Quarteroni, A., Gervasio, P., &amp; Regazzoni, F. (2025) <em>Combining physics-based and data-driven models: advancing the frontiers of research with Scientific Machine Learning</em> <a href="https://arxiv.org/pdf/2501.18708">https://arxiv.org/pdf/2501.18708</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b28d984140e4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/data-reduced-order-models-and-artificial-intelligence-for-digital-twins-b28d984140e4">Data, Reduced Order Models and Artificial Intelligence for Digital Twins</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Neural Operators and Where to Find Them]]></title>
            <link>https://medium.com/sissa-mathlab/neural-operators-and-where-to-find-them-19af4aa9da3e?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/19af4aa9da3e</guid>
            <dc:creator><![CDATA[Dario Coscia]]></dc:creator>
            <pubDate>Thu, 09 Jan 2025 19:36:15 GMT</pubDate>
            <atom:updated>2026-02-03T10:54:17.326Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial Intelligence for the Natural Sciences progress</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4EhJEsVcqOlliERAMpv0nQ.png" /></figure><p>This is the third article of the series <strong>Deep Learning 4 Natural Sciences</strong>. In this article, we will explore the concept of Neural Operators and examine why they are rapidly gaining attention as a promising new approach to fast numerical solvers within the scientific community.</p><h4>Deep Learning 4 Natural Sciences series</h4><blockquote><a href="https://medium.com/@dariocos99/8634cda1022a">How will deep learning be used to speed up physical simulations?</a></blockquote><blockquote><a href="https://medium.com/sissa-mathlab/blending-neural-networks-with-physics-the-physics-informed-neural-network-d681b6b44eb8">Blending Neural Networks with Physics: the Physics Informed Neural Network</a></blockquote><blockquote><a href="https://medium.com/sissa-mathlab/neural-operators-and-where-to-find-them-19af4aa9da3e">Neural Operators and Where to Find Them</a></blockquote><blockquote>PINA, a Python Software for Scientific Machine Learning</blockquote><h4>What is a Neural Operator and Why Do We Need It?</h4><p>In recent years, the volume of available data simulations has increased, particularly in fields like climate science, engineering, and material design opening new frontiers for Deep Learning (DL). Traditional modeling techniques, and even advanced DL methods like PINNs (for a deeper discussion, see “<a href="https://medium.com/sissa-mathlab/blending-neural-networks-with-physics-the-physics-informed-neural-network-d681b6b44eb8">Blending Neural Networks with Physics: the Physics Informed Neural Network</a>”), have struggled to keep pace with the high variability and scale of modern simulations. These challenges have created a pressing need for a new approach — one that can efficiently handle diverse scenarios and does not require prohibitive computational resources. This is where Neural Operators emerge as a game-changer.</p><p>Neural Operators offer a flexible and scalable solution by learning generalized mappings across function spaces (hence don’t require a fixed discretization), allowing them to address new and unseen problems without the need for retraining. For example, in fluid dynamics, where simulations involve complex and highly variable flow patterns, Neural Operators can generalize across different flow configurations, providing accurate results without the time-consuming process of model adjustment for each new situation.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.canva.com%2Fdesign%2FDAGOehQLgdw%2FsY_RP34DheSF-yCD0dHTFQ%2Fview%3Fembed%26meta&amp;display_name=Canva&amp;url=https%3A%2F%2Fwww.canva.com%2Fdesign%2FDAGOehQLgdw%2FsY_RP34DheSF-yCD0dHTFQ%2Fedit%3Futm_content%3DDAGOehQLgdw%26utm_campaign%3Ddesignshare%26utm_medium%3Dlink2%26utm_source%3Dsharebutton&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=canva" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/38129db91d40f9997c01c0fbebc29265/href">https://medium.com/media/38129db91d40f9997c01c0fbebc29265/href</a></iframe><p>Neural Operators are designed to capture the underlying physics of the systems they model, ensuring that simulations remain reliable and precise, even as they scale to more complex and varied datasets. This balance of efficiency and accuracy positions Neural Operators as a transformative tool in scientific computing, capable of meeting the growing demands of data-intensive fields.</p><h4>The Zoo of Neural Operator Architectures</h4><p>Neural Operators encompass a variety of architectures, each with its unique strengths and applications. Unlike traditional methods that depend on specific grid resolutions, neural operators operate independently of discretization, making them adaptable to various resolutions and problem scales.</p><p>At their core Neural Operators transform an input function <em>a </em>to an output function <em>u. </em>It is important to notice that, differently from classical Neural Networks, Neural operators work directly in the functional space, meaning their inputs and outputs are functions, not just vectors. This allows them to handle complex, continuous data more naturally.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*67VXKf1Hg5N2qNbTDqcmpA.png" /><figcaption>The general structure of a neural operator: The input function a is first encoded into an embedding space, and then processed through multiple layers of integral operators and pointwise non-linearities. Finally, a pointwise decoder generates the output function u.</figcaption></figure><p>The general structure of a Neural Operator comprehends three components:</p><ol><li><strong>Encoder: </strong>mapping the input into a specific embedding space. This step is crucial for transforming the input into a format suitable for the neural operator’s kernel.</li><li><strong>Processor:</strong> multiple layers performing function convolution, that is the core computational unit. They process the encoded input to learn convolutional kernels that capture the necessary features and patterns for the task.</li><li><strong>Decoder:</strong> mapping the processors’s output back into the desired output space, translating the processed embedding into the final functional output.</li></ol><p>By varying the design and implementation of these three components — encoder, kernel, and decoder — different neural operators are created, each tailored for specific applications or types of data. For instance: <strong>Fourier Neural Operator (FNO) [1] </strong>utilizes the Fourier transform in the processor to perform global convolutions, capturing long-range dependencies efficiently; <strong>Message Passing Neural Operator (MPNO) [2] </strong>leverages graph neural networks to exchange information between nodes, to perform convolution on unstructured domain; <strong>Universal Transformer Neural Operator (UTNO) [3] </strong>use transformers processors on a fixed embedded space obtained with a graph encoder, offering flexibility and scalability in handling a variety of problem domains.</p><h4>Training Neural Operator Architectures</h4><p>Despite their complex architecture, neural operators are trained using a straightforward, data-driven, supervised approach. This training involves learning from pairs of input and output functions derived from high-fidelity simulations — similar to how traditional classification models are trained on labeled images for tasks like classifying cats and dogs.</p><p>For time-independent systems, a basic loss function, defined as the error between the network’s prediction and the true output function, is typically sufficient. However, for time-dependent systems, more sophisticated methods are needed to accurately predict future states [2].</p><p>In conclusion, as the volume of data generated by simulations continues to grow exponentially, neural operators provide a scalable approach to developing surrogate solvers. These advanced models not only enable the generalization of complex systems but also significantly accelerate the process of obtaining solutions, paving the way for more efficient scientific discoveries. By harnessing the power of neural operators, researchers can tackle previously intractable problems, optimize simulations across various domains, and explore new frontiers in computational science with unprecedented speed and precision.</p><h4>References</h4><ol><li>Li, Zongyi, et al. “Fourier neural operator for parametric partial differential equations.” <em>arXiv preprint arXiv:2010.08895</em> (2020).</li><li>Brandstetter, Johannes, Daniel Worrall, and Max Welling. “Message passing neural PDE solvers.” <em>arXiv preprint arXiv:2202.03376</em> (2022).</li><li>Alkin, Benedikt, et al. “Universal physics transformers.” <em>arXiv preprint arXiv:2402.12365</em> (2024).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19af4aa9da3e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/neural-operators-and-where-to-find-them-19af4aa9da3e">Neural Operators and Where to Find Them</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Challenges in cardiovascular care: a personalized approach]]></title>
            <link>https://medium.com/sissa-mathlab/challenges-in-cardiovascular-care-a-personalized-approach-53cd91eef6a5?source=rss----2639606b2015---4</link>
            <guid isPermaLink="false">https://medium.com/p/53cd91eef6a5</guid>
            <dc:creator><![CDATA[Surabhi Rathore]]></dc:creator>
            <pubDate>Thu, 09 Jan 2025 19:35:11 GMT</pubDate>
            <atom:updated>2025-01-09T19:35:11.786Z</atom:updated>
            <content:encoded><![CDATA[<h4>Overview:</h4><p>Cardiovascular disease remains a leading cause of mortality worldwide. Despite substantial advances in diagnostic imaging, treatments, and devices, there is still a pressing need for more patient-specific, predictive, and adaptive approaches to improve outcomes [1]. Real-time cardiovascular simulations represent a promising avenue for delivering personalized treatment planning, surgical guidance, device design, and continuous patient monitoring. Achieving this, however, is a considerable feat. The complexity and variability of cardiovascular physiology and the computational demands of high-fidelity simulations pose significant technical challenges.</p><h4>Challenges in real-time cardiovascular simulations:</h4><ul><li><em>Acquiring high-quality imaging data</em> (e.g., CT and MRI) and accurately segmenting anatomical structures is crucial.</li><li><a href="https://www.sciencedirect.com/science/article/abs/pii/S0045782505002586"><em>Defining appropriate boundary conditions</em></a><em> [2]</em> is crucial for accurate simulations.</li><li>Complex cardiovascular models demand <em>significant computational resources.</em></li><li>Real-time simulations require <a href="https://link.springer.com/article/10.1007/s13160-021-00469-9"><em>efficient algorithms</em></a><em> [3] </em>and<em> high-performance computing.</em></li><li><em>Computing patient-specific haemodynamics [5] </em>is challenging, especially when tailoring clinical data, such as 4D MRI data.</li></ul><p>By addressing these technical hurdles — <em>improved imaging, </em><a href="https://www.sciencedirect.com/science/article/abs/pii/S0045782505002586"><em>boundary condition specification</em></a><em>, computational efficiency, and </em><a href="https://arxiv.org/pdf/2410.20828"><em>data assimilation</em></a> — researchers and clinicians can advance real-time cardiovascular simulations, ultimately enhancing patient care.</p><h4>State-of-art for real-time cardiovascular simulations:</h4><p>State-of-the-art cardiovascular simulations rely on accurate, patient-specific outflow boundary conditions — often determined via optimal control methods — to achieve realistic flow patterns. As simulations remain computationally costly, reduced-order modelling techniques compress the underlying mathematical complexity — extracting essential features from the high-fidelity model — so that computational costs are dramatically lowered without compromising accuracy. This integration of advanced techniques enables patient-specific haemodynamic predictions, informing personalized medical decisions, such as treatment planning, device design, and surgical interventions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yjCnXyE3skSOGQZbtpe6YA.png" /><figcaption>Optimization Problem Framework for Cardiovascular Modelling</figcaption></figure><h4>Optimal control problems:</h4><p>Outflow boundary conditions play a critical role in determining the accuracy and stability of cardiovascular simulations. Optimal control techniques are employed to refine outflow boundary conditions, ensuring accurate and stable cardiovascular simulations. By treating these boundary conditions as control variables and minimizing a specific objective function, such as the difference between simulated and measured flow velocities or pressures, researchers can achieve more realistic flow <em>(patient-specific measurements obtained from 4D-flow MRI)</em> patterns and enhance the predictive capabilities of cardiovascular flows.</p><p>Optimal control problems for cardiovascular flow involve three key components:</p><ol><li><strong>Forward model:</strong> Solving the <em>unsteady Navier–Stokes equations with suitable conditions</em> to compute haemodynamics.</li><li><strong>Cost functional:</strong> Define an objective that <em>measures the discrepancy between the simulated and target</em> flow fields.</li><li><strong>Lagrange formalism:</strong> Obtain the <em>adjoint equations to compute the gradients of the cost functional</em> concerning control variables.</li></ol><h4>Reduced order modelling:</h4><p>ROM is a technique aimed at approximating high-fidelity, large-scale computational models with significantly fewer degrees of freedom, while still preserving essential accuracy. Extracting the most dominant features from the system’s behaviour is especially valuable in scenarios like parameter studies and real-time simulations.</p><p>Two specific categories of ROM [4] are distinguished:</p><ol><li><strong>Intrusive methods: </strong>These methods begin with the detailed governing equations — <em>like the Navier–Stokes equations for fluid flows</em> — and identify a set of dominant modes from high-fidelity simulations. By projecting the original equations onto this reduced set of modes, we obtain a far-reduced system that still reflects the underlying physics. These models ensure theoretical and physical interpretability but require direct access to modify the original solver and its equations.</li><li><strong>Non-intrusive methods: </strong>Instead of working directly with the full equations, these approaches treat the original simulation code as a <strong>“black box.”</strong> Using data-driven techniques — <em>such as machine learning, regression, or interpolation</em> — the relationship between input parameters and output fields is learned from a collection of precomputed “snapshot” data. These methods are quick to implement since they don’t alter the original solver. However, their accuracy depends on the quality of the training data, and they often provide limited insight into the underlying physics.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3cCkYAVuA8FnMeyeB1G6ug.png" /><figcaption><strong>Comparison of the streamlines for the patient-specifc CABG model [5]</strong></figcaption></figure><p>Carefully selecting and combining these ROM approaches can produce efficient yet accurate reduced solutions suitable for real-time analysis and decision-making.</p><h4>Conclusion:</h4><p>Real-time cardiovascular simulations can revolutionize patient care by overcoming computational, accuracy, and timing barriers. With advances in imaging, optimal control, and reduced-order modelling, these tools support personalized treatments, refined device designs, and precise surgical guidance. As they evolve, clinicians gain the ability to make more informed, personalized decisions, improving outcomes and advancing global health.</p><h4>References:</h4><ol><li>Quarteroni, Alfio, and Luca Formaggia. <em>“Mathematical modelling and numerical simulation of the cardiovascular system.”</em> Handbook of numerical analysis 12 (2004).</li><li>Vignon-Clementel, Irene E., et al. <em>“Outflow boundary conditions for three-dimensional finite element modeling of blood flow and pressure in arteries.”</em> CMAME 195.29–32 (2006).</li><li>Rathore, Surabhi, et al. <em>“Numerical computation of blood flow for a patient-specific hemodialysis shunt model.” </em>JJIAM 38.3 (2021).</li><li>Hesthaven, Jan S., Gianluigi Rozza, and Benjamin Stamm. <em>“Certified reduced basis methods for parametrized partial differential equations.”</em> Vol. 590. Berlin: Springer (2016).</li><li>Rathore, Surabhi, et al. <em>“Projection-based Reduced Order Modelling for Unsteady Parametrized Optimal Control Problems in 3D Cardiovascular Flows.”</em> arXiv preprint arXiv:2410.20828 (2024).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=53cd91eef6a5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sissa-mathlab/challenges-in-cardiovascular-care-a-personalized-approach-53cd91eef6a5">Challenges in cardiovascular care: a personalized approach</a> was originally published in <a href="https://medium.com/sissa-mathlab">SISSA mathLab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>