← Latest papers
🔬 materials science

Physics Informed Priors for Machine Learning Interatomic Potentials with Uncertainty Quantification

This paper introduces a misspecification-aware uncertainty quantification scheme for machine learning interatomic potentials that integrates user-defined physical priors into the model ensemble, ensuring physical consistency while reducing the computational cost of dataset construction for applications like phononic vibrations and elastic properties.

Original author: George I. Simmons, Thomas D. Swinburne, Thomas Hudson, James R. Kermode

Published 2026-11-09
📖 7 min read🧠 Deep dive

Original author: George I. Simmons, Thomas D. Swinburne, Thomas Hudson, James R. Kermode

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 author. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to understand how materials behave, scientists often rely on powerful computer simulations. To run these simulations efficiently, they use simplified models called interatomic potentials. Think of these models as a set of rules that tell a computer how atoms push and pull on one another, allowing researchers to predict everything from how a metal bends to how a crystal melts without needing to solve the most complex equations of quantum physics for every single atom. However, these simplified rules are never perfect. They are approximations, and because they are approximations, they can sometimes lead the simulation astray, predicting that a stable material will suddenly collapse or that an atom will vibrate in a way that physics says is impossible. The challenge for scientists is not just to make a prediction, but to know how much they can trust it, especially when the model itself might be fundamentally flawed in its design.

A team of researchers has developed a new way to fix this problem by teaching these computer models to respect the basic laws of physics while they learn. In a recent study, they introduced a method that acts like a safety net, ensuring that the computer's predictions stay within the realm of physical reality. The researchers focused on a technique called uncertainty quantification, which is essentially a way for a computer model to say, "I am not entirely sure about this answer." Previous methods could tell a model how uncertain it was, but they often failed to stop the model from making physically impossible predictions, such as suggesting that a crystal lattice would vibrate with imaginary speeds or that a solid material would have negative stiffness. The new approach combines the ability to measure uncertainty with strict physical rules, forcing the model to admit its limitations without ever violating the fundamental principles that govern how matter holds together.

The core of this work involves a specific type of machine learning model known as a machine learning interatomic potential. These models are trained on data generated by highly accurate, but extremely slow, quantum mechanical calculations. The goal is to create a fast model that mimics these slow calculations. However, because the fast model cannot capture every detail of the complex quantum world, it inevitably makes mistakes. The researchers used a method called Pointwise Optimal Parameter Set Regression, which generates a whole family of slightly different models rather than just one single "best" model. This family of models helps scientists understand the range of possible outcomes. The problem with this method, until now, was that some of the models in this family might predict that a material is unstable or that its atoms behave in ways that defy the laws of physics. The new study solves this by adding a layer of constraints during the training process. These constraints act as guardrails, ensuring that every single model in the family respects specific physical truths, such as the fact that a crystal must have a stable shape and that its vibrations must be real, measurable quantities.

To test their idea, the researchers applied this method to aluminum, a common metal. They trained their models on a dataset containing thousands of atomic configurations. In one experiment, they focused on the vibrations of the atoms, known as phonons. In a stable crystal, these vibrations must have real frequencies; if a model predicts a vibration with an imaginary frequency, it implies the crystal is falling apart. The researchers found that without their new constraints, the standard method produced a family of models where many members predicted these impossible, imaginary vibrations, suggesting the material was unstable when it should be solid. By enforcing the physical rule that these vibrations must be real, they filtered out the impossible models. The result was a set of predictions that remained uncertain enough to reflect the model's limitations, but were physically consistent. The constrained models correctly predicted that the aluminum crystal would remain stable, whereas the unconstrained models often failed to do so.

The team also looked at how the material responds to being squeezed or stretched, a property described by elastic constants. They trained a more complex neural network model, which is a type of artificial intelligence that learns patterns in data, and applied the same physical constraints. They forced the model to agree with known experimental values for the size of the aluminum crystal and its stiffness, but only within a very small margin of error. This did not force the model to be perfect; instead, it guided the model to stay within a realistic range. When they tested the model on other properties, such as how easily a missing atom (a vacancy) could move through the crystal, they found that the constraints did not ruin the model's ability to predict these new behaviors. The model remained flexible enough to learn the complex details of atomic movement while being anchored to the basic facts of how the material behaves.

The significance of this work lies in its ability to balance two competing needs: the need for a model to admit what it does not know, and the need for it to never say something that is physically impossible. In the past, scientists had to choose between a model that was very uncertain but physically safe, or a model that was precise but might drift into nonsense. This new method allows them to have both. By building physical laws directly into the way the model learns, the researchers ensure that the uncertainty they measure is meaningful. It tells them how much the model's design might be missing, rather than just how much the data is noisy. This is particularly important for simulating materials at different temperatures or under different pressures, where small errors can lead to large, incorrect conclusions about whether a material will melt or break.

The researchers demonstrated that their approach works for different types of models, from simpler linear equations to complex neural networks. They showed that even when the training data was limited, the method could produce reliable results by leaning on the physical constraints. For instance, when they simulated aluminum at a temperature of 200 Kelvin, the unconstrained models predicted a chaotic structure that did not resemble a crystal. The constrained models, however, maintained the correct crystal structure, showing clear peaks in the arrangement of atoms that matched what is expected for a solid metal. This suggests that the method can rescue simulations that would otherwise fail, allowing scientists to run longer and more complex experiments without worrying that the model has drifted into a physical impossibility.

Ultimately, this research offers a practical tool for scientists who design materials. It provides a way to incorporate known facts, such as the size of a crystal or the speed of sound through a metal, directly into the learning process of a computer model. This reduces the need to generate massive amounts of new data just to fix a model's basic errors. Instead of constantly feeding the computer more examples of how a material should behave, scientists can simply tell the computer the rules it must follow. This makes the process of creating accurate models faster and more efficient. The method is designed to be flexible, meaning it can be applied to many different materials and properties, as long as those properties can be described by a set of mathematical rules that the computer can check.

The study concludes that while no model can perfectly replicate the complexity of the quantum world, we can build models that are honest about their limitations without sacrificing physical reality. By enforcing these physical priors, the researchers have created a framework where uncertainty is not just a measure of doubt, but a measure of how much the model's design differs from the true physics, all while keeping the predictions grounded in reality. This approach opens the door to more reliable simulations of complex processes, such as how materials deform under stress or how they respond to extreme heat, giving scientists greater confidence in the results they use to design the materials of the future.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →