🔬 materials science

Thermodynamics-Informed Machine Learning for Energy Materials Discovery

This perspective argues that the next frontier in energy materials discovery lies in developing thermodynamics-informed machine learning models that move beyond static zero-temperature energy predictions to capture the essential roles of entropy and anharmonicity in governing materials behavior under realistic operating conditions.

Original authors: Pol Benítez, Cibrán López, Claudio Cazorla

Published 2026-07-30

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Pol Benítez, Cibrán López, Claudio Cazorla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather. If you only look at a frozen snapshot of the atmosphere at absolute zero—where everything is still and silent—you might think a cloud is just a fluffy, static ball of water. But in the real world, heat makes air dance, winds swirl, and clouds shift shape. In the world of materials science, scientists use powerful computers to design new materials for things like solar panels, batteries, and catalysts. For a long time, they've relied on a "frozen snapshot" approach, calculating how atoms sit still in a perfect, cold grid. This works well for some things, but it misses the most important part of reality: heat. Heat isn't just a warm blanket; it's a chaotic party that changes how atoms move, how they stick together, and even which shapes they prefer to be in. Without accounting for this thermal dance, our computer predictions often fail to match what happens in real life, where everything is hot and moving.

This paper argues that the next big leap in discovering new energy materials requires us to stop looking at frozen snapshots and start teaching our computers to understand the "thermodynamics" of heat. The authors, Pol Benítez, Cibrán López, and Claudio Cazorla, explain that current machine learning models are like students who have only studied a textbook of still images; they are brilliant at predicting the energy of a frozen structure but terrible at guessing how that structure will behave when it's hot and wiggling. They point out that for many materials, the "winner" isn't the one with the lowest energy at zero degrees, but the one that can handle the chaos of heat the best. The paper doesn't claim to have solved this problem yet; instead, it maps out a roadmap for how we can build a new generation of AI that learns not just static energy, but the messy, beautiful, and essential role of temperature in making materials work.

The Frozen Trap vs. The Hot Reality

Think of a material as a Lego castle. In the old way of doing things, scientists used computers to find the most stable castle by looking at how the blocks fit together when everything was perfectly still and cold. This is called the "zero-temperature" approximation. It's like judging a dance competition by only looking at the dancers when the music stops. The problem is, in the real world, the music never stops. The atoms are constantly vibrating, jiggling, and swapping places. This movement creates "entropy," which is a fancy word for the amount of disorder or chaos in a system.

The paper explains that at high temperatures, this chaos can actually make a structure more stable. Imagine a wobbly tower of blocks that looks like it should fall over when it's still. But if you shake the table (add heat), the blocks might jiggle into a new, tighter formation that holds together better than the still version. Current machine learning models often miss this entirely. They are trained on data from static, frozen calculations, so they think the wobbly tower is unstable and discard it. But in reality, that wobbly tower might be the champion at room temperature. The authors argue that relying on these frozen models is a major barrier because it leads us to pick materials that fail as soon as we turn on the heat.

The New Game Plan: Teaching AI to Feel the Heat

So, how do we fix this? The paper suggests we need to upgrade our AI from a "static photographer" to a "thermodynamic director." Instead of just predicting the energy of a single, still structure, the new models need to predict the "free energy," which is a combination of the static energy plus the effects of heat and entropy.

The authors review a few different strategies that are already being tried, like using "machine-learned interatomic potentials." You can think of these as super-smart shortcuts. Instead of doing the incredibly slow and expensive math for every single atom movement (which is like trying to count every grain of sand on a beach one by one), these AI models learn the rules of the sand's movement and can simulate the whole beach in a flash. This allows scientists to run "molecular dynamics" simulations, which are like virtual movies of atoms dancing at high temperatures. From these movies, we can extract the free energy and see which materials are actually stable.

However, the paper warns that even these advanced tools have limits. If the AI is only trained on calm, low-temperature dances, it might get confused when the music speeds up and the atoms start jumping wildly. The authors suggest a few key fixes:

Direct Learning: Instead of simulating the dance and then calculating the energy afterward, we should teach the AI to predict the free energy directly, treating temperature as a variable it can learn, just like it learns the shape of an atom.
Entropy Awareness: The AI needs to understand that entropy (the chaos factor) is a real thing that changes the rules. We need to build models that can "feel" the difference between a quiet room and a crowded party.
Active Learning: This is like a smart student who knows exactly what they don't understand. Instead of studying random facts, the AI asks, "Hey, I'm not sure about this specific temperature range; let's run a simulation there to learn more." This helps fill in the gaps where the data is missing.

Why This Matters for Energy

The paper highlights three specific areas where this new approach could be a game-changer:

Solar Panels (Photovoltaics): In materials like perovskites, the heat can change the color of light the material absorbs. If we only look at the frozen version, we might think a material is useless, but when it's hot, it might be perfect for capturing sunlight. The heat can also change how easily electricity flows through the material.
Catalysts: These are materials that speed up chemical reactions, like in car exhaust systems or making fuel. The surface of a catalyst isn't a rigid rock; it's a dynamic landscape that changes shape when it gets hot. The "active sites" where the magic happens might only appear when the atoms are jiggling. If we design a catalyst based on a frozen model, we might miss the real hero.
Solid-State Batteries: These batteries use solid materials to move ions (charged atoms) instead of liquid. The movement of these ions is often helped by the vibrations of the material's structure. If the structure is too rigid (frozen), the ions can't move. But if the structure is "soft" and wiggly at the right temperature, the ions can zip through. Current models often miss this "softness" because they don't account for the heat-induced vibrations.

The Road Ahead

The authors are clear that we aren't there yet. They admit that getting the data to train these new models is hard and expensive. Simulating heat requires a lot of computer power, and we don't have enough examples of materials behaving at different temperatures to teach the AI everything it needs to know. They also point out that current AI models often break the rules of thermodynamics, predicting things that are physically impossible, like negative heat capacity.

However, the paper is optimistic. It suggests that by combining the speed of machine learning with the strict rules of thermodynamics, we can build a new kind of AI that doesn't just guess what a material looks like, but understands what it does when it's working. The goal is to move from designing materials for a frozen world to designing them for the real, hot, chaotic world where they will actually be used. It's a shift from taking a still photo to directing a full-motion movie, and it promises to unlock a new era of energy materials that are more efficient, stable, and powerful than anything we could design before.

Technical Summary: Thermodynamics-Informed Machine Learning for Energy Materials Discovery

Problem Statement
Current machine learning (ML) models for materials discovery are fundamentally limited by their reliance on zero-temperature ( $T=0$ K) descriptions. These models are typically trained on static lattice energies derived from Density Functional Theory (DFT) calculations, learning a mapping from atomic configuration to internal energy ( $U$ ) alone. However, material stability, phase transitions, and functional performance under realistic operating conditions are governed by the Gibbs free energy ($G = U - TS + PV$), where entropic contributions ( $S$ ) from lattice vibrations, configurational disorder, electronic excitations, and magnetic fluctuations play a decisive role.

The zero-temperature approximation leads to qualitative failures in predicting experimentally observed behavior, particularly in systems with strong anharmonicity (e.g., halide perovskites) or competing polymorphs. A phase that is dynamically unstable at $T=0$ K may become thermodynamically stable at finite temperatures due to entropic stabilization. Consequently, ML-accelerated screening pipelines based on static convex hulls often misidentify stable phases, miss thermally stabilized structures, and fail to predict temperature-dependent properties such as band-gap renormalization, ionic conductivity, and catalytic activity.

Methodology and Current Strategies
The paper reviews existing strategies that attempt to bridge the gap between static ML and finite-temperature thermodynamics, categorizing them into three main approaches:

Machine-Learned Interatomic Potentials (MLIPs): These models regress the potential energy surface (PES) onto flexible functional forms trained on first-principles data. Once trained, they enable molecular dynamics (MD) simulations at a fraction of the cost of ab initio MD. Thermodynamic quantities (free energy, heat capacity) are subsequently extracted via thermodynamic integration or free-energy perturbation. While foundational potentials (e.g., MACE, CHGNet) offer broad transferability, they often lack sufficient training data for large-amplitude anharmonic displacements and phase transitions.
Thermodynamic Workflows with ML Surrogates: This approach retains established statistical mechanics formalisms (e.g., Quasi-Harmonic Approximation, Self-Consistent Phonon theory) but uses ML to accelerate expensive steps, such as predicting force constants or phonon spectra. This ensures thermodynamic consistency by construction but inherits the physical limitations of the underlying approximations (e.g., failure in strongly anharmonic systems).
Direct Learning of Temperature-Dependent Properties: This emerging strategy treats temperature-dependent observables (e.g., phonon dispersions, thermal conductivity, renormalized band gaps) as primary regression targets, learning a direct map from structure (and sometimes temperature) to the property. While efficient, these models often suffer from data scarcity, limited generalization across phase transitions, and a lack of guaranteed thermodynamic consistency between predicted quantities (e.g., free energy and entropy).

Key Contributions and Proposed Roadmap
The authors argue that the field must shift from predicting static energies to modeling free energies directly. They propose a roadmap for "thermodynamics-informed ML" centered on three pillars:

Direct Free-Energy Learning: Developing models that take temperature as an explicit, learnable input to predict free energy ( $F$ ) directly. This involves learning joint representations of structure and temperature, potentially using multi-task learning to simultaneously predict free energy, entropy ( $S = -\partial F/\partial T$ ), and heat capacity ( $C_v = -T\partial^2 F/\partial T^2$ ) to enforce thermodynamic consistency. Architectures must ensure smoothness and differentiability to avoid unphysical oscillations in derived quantities.
Entropy-Aware Representations: Moving beyond total free energy to explicitly model entropic contributions. The paper highlights information-theoretic approaches (e.g., MICE) that decompose entropy into mutual information terms, allowing for the estimation of configurational and vibrational entropy without exhaustive sampling.
Active Learning in Temperature Space: Extending active learning loops to navigate the composition-temperature phase space. Instead of filtering candidates based on zero-temperature convex hulls, workflows should utilize finite-temperature free-energy hulls. This involves iteratively selecting high-uncertainty points in both composition and temperature for high-fidelity evaluation, ensuring the model captures the shifting stability landscapes near phase transitions.

Results and Applications in Energy Materials
The paper illustrates the necessity of this paradigm through three specific energy material families where temperature is constitutive of the target property:

Photovoltaic Absorbers: In materials like halide perovskites, electron-phonon coupling (EPC) and lattice anharmonicity cause significant band-gap renormalization (up to 60% reduction at room temperature) and limit carrier mobility. Static models fail to capture these effects. Recent ML work (e.g., HamEPC, non-perturbative frozen-phonon schemes) demonstrates that learning EPC and temperature-dependent band structures is essential for accurate device performance prediction.
Heterogeneous Catalysts: Catalytic activity is governed by Gibbs free energy barriers, not static potential energy barriers. Anharmonic effects can lower activation barriers by 40% or even reverse the rate-limiting step. Furthermore, catalyst surfaces are dynamic; adsorbate-induced reconstructions and thermal fluctuations create active sites absent in static slab models. MLIPs combined with umbrella sampling have successfully revealed these dynamic mechanisms.
Solid-State Electrolytes (SSE): Ionic conductivity relies on migration entropy and the anharmonic renormalization of migration barriers. Many SSEs are metastable at $T=0$ K but become stable superionic conductors at finite temperatures due to order-disorder transitions. MLIPs trained on AIMD trajectories have enabled the simulation of long-timescale diffusion and the accurate prediction of Arrhenius activation energies, which static NEB calculations often underestimate by up to 1 eV.

Significance and Claims
The paper claims that developing thermodynamics-informed ML constitutes one of the most important and least explored frontiers in materials discovery. It asserts that the zero-temperature approximation is not merely a technical shortcoming but a fundamental barrier to predictive design under realistic conditions.

The significance of this perspective lies in its call for a conceptual reorientation:

From Static to Thermodynamic: The target quantity must shift from internal energy ( $U$ ) to free energy ( $F$ ).
From Post-Hoc to Explicit: Temperature must be treated as an explicit, learnable variable within the model architecture, rather than a post-hoc quantity derived from expensive sampling.
From Accuracy to Consistency: Future models must prioritize thermodynamic consistency (ensuring derived quantities like entropy and heat capacity obey physical laws) alongside predictive accuracy.

The authors conclude that achieving this requires a deeper integration of ML with thermodynamic principles, the generation of high-quality finite-temperature training data, and the development of foundational models capable of transferring across chemical space while capturing anharmonic and entropic effects. This paradigm shift promises to transform computational materials design from predicting what a material is at absolute zero to predicting what it does at the temperatures where it must actually work.

The Frozen Trap vs. The Hot Reality

The New Game Plan: Teaching AI to Feel the Heat

Why This Matters for Energy

The Road Ahead

More like this