<?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[Stories by Gracie Chaney on Medium]]></title>
        <description><![CDATA[Stories by Gracie Chaney on Medium]]></description>
        <link>https://medium.com/@graciechaney21?source=rss-1aeca46e0b16------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*yiKtg7phc-M2zl3h.jpg</url>
            <title>Stories by Gracie Chaney on Medium</title>
            <link>https://medium.com/@graciechaney21?source=rss-1aeca46e0b16------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 23:06:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@graciechaney21/feed" 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 Machine Learned Interatomic Potentials (MLIPs)]]></title>
            <link>https://medium.com/@graciechaney21/introduction-to-machine-learned-interatomic-potentials-mlips-68c4d2f17874?source=rss-1aeca46e0b16------2</link>
            <guid isPermaLink="false">https://medium.com/p/68c4d2f17874</guid>
            <category><![CDATA[chemistry]]></category>
            <category><![CDATA[density-functional-theory]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[molecular-dynamics]]></category>
            <category><![CDATA[physics]]></category>
            <dc:creator><![CDATA[Gracie Chaney]]></dc:creator>
            <pubDate>Tue, 22 Apr 2025 19:49:11 GMT</pubDate>
            <atom:updated>2025-04-22T19:49:11.929Z</atom:updated>
            <content:encoded><![CDATA[<p>With all the hype about ChatGPT and stable diffusion, it is easy to miss the more niche AI developments. Various scientific fields have been incorporating AI into their research, from using computer vision to detect per-cancerous breast tissue¹ to using generative AI to create first guesses of potential drug molecules.² It is not surprising that entire sub-fields of research have been developed to better utilize and alter AI algorithms for specific uses. Rather than AI simply being a tool for research, it has become an object of research itself.</p><p>For the past 2.5 years, I have been working as a post-doctoral researcher to simulate material interactions at an atomic level. Previously, I had worked only on <em>ab initio </em>simulations that — although highly accurate — are constrained to a small number of atoms and unrealistically short time-scales. Nowadays, I am able to use these simulations as data for a training-set for a neural network that trains off of the energies and forces of the atoms’ local environments. What we end up with is a map of the potential energy landscape (Fig 1³)of a material that enables us to predict how the material will behave over extended periods of time and at various temperatures and pressures. Keep reading to learn how.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*jenFuuVNgO2V18rZlKzjEQ.png" /><figcaption>Fig 1: Image taken from S. Jin et al.³ The potential energy landscape describes the energy of interacting atoms. Useful concept in material and biological sciences.</figcaption></figure><p><strong>But what’s wrong with what we had before?</strong></p><p>It is not possible to do <em>ab initio</em> calculations justice in a single article, let alone a paragraph. In VERY broad strokes, <em>ab initio </em>(also known as Density Functional Theory) calculations are based on the laws of quantum mechanics. The goal is to estimate the solution to the many-body Schrodinger equation. As the “bodies” in question are atoms and molecules, the main idea is that we want to be able to predict the interactions between all of the atoms in a system given <strong>just the atomic positions — </strong>hence the name, <em>ab initio</em>. If this sounds hard, you’re right! We can make this problem simpler by assuming that the nucleus of all the atoms are static compared to the electrons and can be ignored in the calculation. We can also choose to focus on the electron <strong>density </strong>rather than the individual electrons themselves. With these adjustments, we are able to predict potential energy surface (PES) and atomic forces for anything from individual drug molecules to continuous materials like graphene. These predicted PES and sets of forces allows us to run something called <em>ab initio</em> molecular dynamics (AIMD) to study dynamical systems that consist of moving atoms (even just from thermal fluctuations). This method is powerful, and enables us to study the evolution of systems at quantum accuracy. So why aren’t we satisfied?</p><p>The major drawback of <em>ab initio </em>calculations is their computational inefficiency. We are generally constrained to systems with less than 1000 atoms and are limited to a few-hundred picoseconds (10^-12 seconds). There is little point in running simulations that are extremely accurate for systems/events that are too small/short lived to be realistic. On the other hand, we can choose to ignore quantum effects entirely, and perform only classical molecular dynamics; however, what we gain in performance, we lose in accuracy. Another solution, is to <strong>bridge-the-gap between the accuracy of the quantum simulations and the efficiency of the classical ones.</strong></p><p><strong>The Solution</strong></p><p>The <em>ab initio</em> calculations provide us with highly accurate interatomic energy and forces that can be used as target values to train a machine learning model known as a machine-learned interatomic potential (MLIP). There are many different flavors of MLIPs, some based on linear regression and others on neural networks architecture. <strong>A key similarity between all of them is their use of “structural descriptors” to encode information about the local atomic environments </strong>in a way that the algorithm can understand (so normally a matrix of some dimension). There are various rules that determine what can and what cannot be a structural descriptor.⁴ For instance, an atomic environment does not change just because it’s moved around in space or rotated, so neither should its descriptor! This is the reason why structural descriptors are also called “symmetry functions.”</p><p>One of the first MLIPs, known as <strong>the Behler-Parrinello neural network</strong>,⁵ uses two types of descriptors: the “radial descriptor” that includes the radial distances between all of the atom-pairs, and the “angular descriptor” that includes the angles between all triples of atoms. These descriptors are are centered around a particular atom and “see” only the atoms within a certain radial cutoff of that atom, as shown in Fig 1. This radial cutoff is on the order of angstroms (10^⁻10 m) and is generally chosen by the user (aka. me). Increasing the value improves the accuracy, while worsening computational efficiency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/569/1*WQCksO7DLmtEF3mtCqTJcw.png" /><figcaption>Fig 1: Basic idea of Behler-Parrinello descriptor with both radial and angular components. The local environment contains all of the atoms (light blue) within a radial cutoff “R.” The atoms outside of this region are drawn as dark, as the functions ignore them for the time being.</figcaption></figure><p>There are many other flavors of MLIP besides Behler-Parrinello; however, this article is just meant to be a brief introduction.</p><p>Above all, the most important thing to note is that MLIPs only require the <em>ab initio </em>training data, a type of structural descriptor, and a machine learning algorthm. <strong>They do NOT require any experimental data! </strong>Although researchers in the past have developed interatomic potentials from empirical means, they require hand-tuned parameters that can be specific to the system they are working on. Thus, developing an empirical potential is very time consuming and requires a great deal of expertise. Training and use of MLIPs have a much easier barrier of entry, and although the finished models themselves are not transferable to all systems, the base architecture remains the same.</p><p><strong>What have we been able to accomplish with MLIPs?</strong></p><p>In the past few years, the computational science community has used MLIPs to further research into renewable energy solutions and the prediction of future drugs. Since this article is getting rather long, let us just focus on one example in battery design.</p><p>Discovering better materials for energy-dense batteries is a hot-topic in material science (and is probably necessary to our continued survival on this planet). The main parts of a battery are the cathode (positive side), anode (negative side) and the electrolyte in between. Not every combination of cathode, anode and electrolyte works, and it can be difficult to know from intuition before testing. As experiments are costly in both time and money, it makes sense to screen these battery materials beforehand with some (relatively) cheap computations. In a 2024 paper by Choyal <em>et al.,⁶ </em>researchers in India trained a few MLIPs off of a dataset of lithium-based disordered rocksalts (LDRs). The important thing to note about LDRs is that they are potential cathode materials that could increase the energy density of Li-ion batteries. Fig 2 illustrates the team’s workflow:</p><ol><li>The <em>ab initio</em> (DFT) datasets, containing ultimatly 10842 structures of LDRs and their associated energies and forces, were generated and cleaned.</li><li>Four flavors of MLIP were trained on this dataset.</li><li>The MLIPs were used to obtain the predicted energies for various LDR structures.</li><li>With these energies, the predicted voltage was calculated and compared to the “true” DFT voltage, to test the accuracy of the four MLIP types (note: accuracy of a computational tool is relative to the physical calculation and to the material).</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/790/1*EHJJxYx3LsdsamUCcew6mw.png" /><figcaption>Fig 2: From Choyal <em>et al.⁶ Workflow of DFT (ab initio) data generation, MLIP training, and final comparison of results.</em></figcaption></figure><p>[1]: A. Carriero, L. Groenhoff, E. Vologina, P. Basile, and M. Albera. “Deep Learning in Breast Cancer Imaging: State of the Art and Recent Advancements in Early 2024.” <em>Diagnostics (Basel, Switzerland)</em>, <em>14</em>(8), 848. (2024). <a href="https://doi.org/10.3390/diagnostics14080848">https://doi.org/10.3390/diagnostics14080848</a></p><p>[2]: H. H. Loeffler, J. He, A. Tibo <em>et al.</em> “Reinvent 4: Modern AI–driven generative molecule design.” <em>J Cheminform</em><strong> </strong>(16) 20 (2024). <a href="https://doi.org/10.1186/s13321-024-00812-5">https://doi.org/10.1186/s13321-024-00812-5</a></p><p>[3] S. Jin, X. Fan, C. Stamper <em>et al.</em>“On the temperature dependence of the density of states of liquids at low energies.” <em>Scientific Reports.</em> 14. (2024) 10.1038/s41598–024–69504–2.</p><p>[4]: G. Wang, C. Wang, X. Zhang <em>et al.</em> “Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations.” <em>iScience.</em> (27)5. (2024). <a href="https://doi.org/10.1016/j.isci.2024.109673">https://doi.org/10.1016/j.isci.2024.109673</a>.</p><p>[5]: J. Behler and M. Parrinello. “Generalized neural-network representation of high-dimensional potential-energy surfaces.” <em>Phys. Rev. Lett.</em>, (2007). <a href="https://doi.org/10.1103/physrevlett.98.146401">doi:10.1103/physrevlett.98.146401</a>.</p><p>[6]: V. Choyal, N. Sagar, and G. Sai Gautam. “Constructing and Evaluating Machine-Learned Interatomic Potentials for Li-Based Disordered Rocksalts.” <em>Journal of Chemical Theory and Computation</em> <em>(20)</em>11, 4844–4856 (2024). <a href="http://dx.doi.org/10.1021/acs.jctc.4c00039">http://dx.doi.org/10.1021/acs.jctc.4c00039</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=68c4d2f17874" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>