<?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 JengYuan Chia on Medium]]></title>
        <description><![CDATA[Stories by JengYuan Chia on Medium]]></description>
        <link>https://medium.com/@jengyuanchia?source=rss-1dd5db417b1a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*-yuQiuqEchU2btHW.</url>
            <title>Stories by JengYuan Chia on Medium</title>
            <link>https://medium.com/@jengyuanchia?source=rss-1dd5db417b1a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 18:27:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jengyuanchia/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[Why Physics-Informed AI Will Matter More Than Bigger AI Models.]]></title>
            <link>https://medium.com/@jengyuanchia/why-physics-informed-ai-will-matter-more-than-bigger-ai-models-9af0778ea96a?source=rss-1dd5db417b1a------2</link>
            <guid isPermaLink="false">https://medium.com/p/9af0778ea96a</guid>
            <dc:creator><![CDATA[JengYuan Chia]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 15:22:55 GMT</pubDate>
            <atom:updated>2026-08-02T15:22:55.986Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>The next breakthrough in industrial AI may not come from adding more parameters, but from teaching machines how the physical world actually works.</blockquote><h4>For much of the past decade, progress in artificial intelligence has followed a relatively simple formula:</h4><p><strong>More data. More computing power. Bigger models.</strong></p><p>The results have been remarkable.</p><p>Modern AI can generate convincing text, interpret images, write software, translate languages and solve increasingly complex problems. Every new generation of models appears to be more capable than the previous one.</p><p>This has created a widespread assumption that scaling alone may eventually solve most AI problems.</p><p>But there is an important question that receives far less attention:</p><p><strong><em>Does a bigger model necessarily understand the physical world better?</em></strong></p><p>Not always.</p><p>A model may recognize millions of visual patterns without understanding how light forms an image. It may predict equipment behavior without understanding force, heat, vibration, energy or material deformation.</p><p>It may produce an answer that is statistically convincing while still violating the physical rules of the system it is supposed to represent.</p><p>For conversational AI or image generation, such errors may occasionally be tolerated.</p><p>For industrial inspection, autonomous systems, medical devices, energy systems or safety-critical engineering, they cannot.</p><p><strong>In the physical world, being statistically convincing is not enough.</strong></p><p>The prediction must also be physically plausible, reliable under changing conditions, consistent with engineering knowledge and safe enough to support real decisions.</p><p><strong>This is where physics-informed AI becomes important.</strong></p><h3>Bigger Models Do Not Automatically Produce Physical Understanding</h3><p>Large AI models are extremely good at learning correlations.</p><p>Given enough examples, they can detect patterns that would be difficult for humans to describe explicitly. This ability has already transformed computer vision, natural-language processing and scientific computing.</p><p>However, correlation is not the same as physical understanding.</p><p>Consider an AI vision system inspecting automotive components.</p><p>Under the lighting conditions represented in its training dataset, the model may achieve excellent accuracy. But once the illumination angle changes, reflections appear, the camera exposure drifts or the surface material varies slightly, performance may deteriorate.</p><p>The component itself may not have changed.</p><p>What changed was the physics of image formation.</p><p>A conventional data-driven model may interpret the new appearance as an unfamiliar pattern. It does not inherently know whether the variation came from a genuine defect, changing illumination, camera response, geometry or surface reflectance.</p><p>A physics-informed approach attempts to include some of that knowledge within the AI system.</p><p>The difference is important:</p><p><strong><em>A purely data-driven model learns what the training images look like. A physics-informed model also considers why the images look that way.</em></strong></p><p>That does not mean the AI possesses human-like physical understanding. It means the learning process is guided by knowledge that engineers already possess.</p><h3>What Does “Physics-Informed” Actually Mean?</h3><p>The term physics-informed AI is sometimes associated mainly with Physics-Informed Neural Networks, or PINNs.</p><p>PINNs incorporate governing equations, such as ordinary or partial differential equations, into the model’s training objective. Instead of learning only from labelled data, the network is also penalized when its predictions violate the governing physical equations.</p><p>For example, a neural network modelling temperature may be required to satisfy a heat-transfer equation. A fluid model may be constrained by conservation of mass and momentum. The governing equation becomes part of the learning process rather than merely something checked after training.</p><p>But physics-informed AI is broader than PINNs.</p><p>· Physics can be introduced through governing equations, boundary and initial conditions, conservation laws, physically meaningful model architectures, simulation-generated data, optical or mechanical models, constraints on impossible predictions and preprocessing based on known physical behavior.</p><p>The central idea is simple:</p><p><strong><em>The model should not be forced to rediscover every law of the physical world from data alone.</em></strong></p><p>When reliable prior knowledge already exists, ignoring it may be wasteful.</p><figure><img alt="Figure 1. The progression from data-driven AI to physics-informed AI and ultimately to more reliable engineering intelligence." src="https://cdn-images-1.medium.com/max/1024/1*_0l0nH0z_SE0lqxUhy7XIg.png" /><figcaption><strong>Figure 1. The progression from data-driven AI to physics-informed AI and ultimately to more reliable engineering intelligence.</strong></figcaption></figure><h3>Data Alone Is Often Not Enough in Industry</h3><p>Many successful AI systems are built using huge datasets.</p><p>Industrial engineering rarely offers that luxury.</p><p>A factory may produce thousands of acceptable components but very few defective ones. Certain defects may appear only occasionally. Production conditions may change across shifts, suppliers, machines or product variants.</p><p>Collecting and labelling enough examples of every possible failure mode may be expensive or impossible.</p><p>· This creates highly imbalanced datasets, insufficient examples of rare defects, changes in lighting and camera positioning, variation in material appearance, new products with little historical data and limited tolerance for false negatives.</p><p>A large model cannot create missing physical information simply by having more parameters.</p><p>In some cases, making a model larger may even increase the risk of learning irrelevant visual correlations from a small dataset.</p><p>Physics-informed methods provide another source of knowledge.</p><p>For example, an industrial vision system may use information about the expected geometry of a component, the direction of illumination, the camera projection model, surface reflection, allowable deformation, expected temperature behavior or mechanical limits.</p><p>Such information does not eliminate the need for data. It helps the model use limited data more intelligently.</p><h3>Physics as a Form of Intelligent Constraint</h3><p>In ordinary machine learning, a model is often allowed to explore a very large space of possible solutions.</p><p>Physics reduces that space.</p><p>Suppose an AI predicts the future motion of an object. Without physical constraints, many mathematically possible trajectories may be considered. But if the model must also obey known limits on velocity, acceleration, force or energy, many impossible solutions can immediately be rejected.</p><p>This is one of the strongest advantages of physics-informed learning.</p><p>Physics acts as an intelligent constraint.</p><p>· It tells the model which outcomes are possible, which behaviors are unlikely, which relationships must remain consistent and which predictions should be rejected even when they fit limited training data.</p><p>This can improve generalization because the model is not learning from examples in isolation. It is learning within a structured description of reality.</p><p>In engineering, constraints are not necessarily weaknesses.</p><p>Good constraints often make intelligent systems more reliable.</p><h3>Industrial AI Needs More Than Benchmark Accuracy</h3><p>AI research frequently focuses on benchmark performance.</p><p>A model may achieve higher accuracy, better mean average precision or lower prediction error than another model. These measures are useful, but they do not fully capture industrial reliability.</p><p>· A factory needs to know whether the model will still work when illumination changes, whether it can distinguish a real defect from glare, what happens when the camera is slightly displaced, whether it will recognize a previously unseen variation, whether engineers can understand why it made a decision and whether the prediction violates known physical behavior.</p><p>An additional one percent improvement on a benchmark may matter less than stable performance under real production variation.</p><p>This changes the definition of a good AI system.</p><p>The best industrial model may not be the largest one.</p><p>It may be the model that combines sufficient learning capacity, appropriate physical knowledge, robust sensing, good engineering judgement and clearly understood failure limits.</p><p>In other words, useful industrial intelligence is not created by AI alone.</p><p><strong><em>It emerges from the integration of data, physics and engineering experience.</em></strong></p><h3>The Future Is Likely to Be Hybrid</h3><p>Physics-informed AI should not be viewed as a rejection of data-driven learning.</p><p>It is a way of strengthening it.</p><p>Purely physics-based models can be limited when real systems are too complex, uncertain or computationally expensive. Purely data-driven models can struggle when data are scarce, operating conditions change or physical consistency matters.</p><p>A hybrid approach can combine the strengths of both.</p><p>Data-driven learning can identify patterns that are difficult to model analytically. Physics can provide structure, constraints and interpretability. Human engineers can define the relevant operating conditions, risks and acceptable limits.</p><p>This creates a more complete form of intelligence:</p><p><strong><em>Data shows the model what has happened. Physics helps explain what can happen. Engineering judgement determines what matters.</em></strong></p><p>· This hybrid approach may become especially important in advanced manufacturing, robotics, energy systems, digital twins, autonomous machines, climate modelling, biomedical engineering, materials science and computational imaging.</p><p>These fields operate in the physical world, where predictions must eventually confront reality.</p><h3>From Pattern Recognition to Engineering Intelligence</h3><p>Most current AI is exceptionally powerful pattern recognition.</p><p>The next stage may require something more.</p><p>AI systems will need to reason within physical environments, work with incomplete data, recognize causal relationships and remain reliable when operating conditions change.</p><p>That transition could move AI from pattern recognition toward what might be called engineering intelligence.</p><p>Engineering intelligence does not require a machine to possess complete human-like understanding. It requires the system to combine learned patterns with meaningful knowledge about the physical process being observed.</p><p>For an industrial vision system, this may involve understanding how illumination, optics, geometry and surface properties interact.</p><p>For a predictive-maintenance system, it may involve relationships between load, vibration, temperature and wear.</p><p>For an energy system, it may involve conservation laws and operating constraints.</p><p>In every case, the AI becomes more useful when its predictions are grounded in the mechanism of the system rather than in correlations alone.</p><h3>Bigger AI Will Still Matter</h3><p>It would be misleading to claim that model scale no longer matters.</p><p>Larger models will continue to produce major advances. More capable foundation models, improved computing hardware and larger datasets will remain important.</p><p>The real argument is not physics instead of scale.</p><p>It is that scale alone is unlikely to be sufficient for every physical and industrial problem.</p><p>The future may belong to systems that combine powerful general models with specialized physical understanding.</p><p>A large model can provide broad capability.</p><p>Physics-informed learning can provide domain grounding.</p><p>Engineering knowledge can provide purpose and practical limits.</p><p>Together, they may create systems that are not only more capable, but also more trustworthy.</p><h3>Conclusion</h3><p>The history of AI has recently been dominated by scale.</p><p>But the next major step in industrial AI may not come only from increasing the number of parameters, GPUs or training examples.</p><p>It may come from teaching AI to respect the structure of the physical world.</p><p>Physics-informed AI offers a path toward models that can learn from limited data, remain physically consistent and perform more reliably under real operating conditions.</p><p>This does not make data-driven AI obsolete.</p><p>It makes it more grounded.</p><p>The most valuable future systems may therefore not be the largest models in isolation. They may be hybrid systems that combine machine learning, physical knowledge and human engineering judgement.</p><p><strong><em>The next stage of AI will not only be about models that recognize more patterns. It will be about models that better understand the physical rules behind those patterns.</em></strong></p><p>And for industrial AI, that difference may matter more than size alone.</p><p><em>About the author</em></p><p><em>Engineer and PhD researcher in Applied Physics. Interested in physics-informed AI, industrial computer vision and practical machine learning.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9af0778ea96a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>