Physics‑Informed GP Model Enables Near‑Infinite Stability in Hot Molecular Dynamics

Machine‑learning potentials (MLPs) are celebrated as the "holy grail" for combining quantum‑level accuracy with molecular‑mechanics efficiency, and over the past two decades they have reproduced static energy and force tests with unprecedented precision. However, many MLPs that excel in static benchmarks explode or collapse when deployed in molecular dynamics (MD) simulations, especially at elevated temperatures, severely limiting their practical use.

Physics‑Informed Gaussian Process Regression Model

A research team from the University of Manchester tackled this stability problem by building a physics‑informed Gaussian Process Regression (GPR) atomic energy model within the Interaction Quantum Atoms (IQA) framework. This model is the first to achieve almost unlimited stability in NVT simulations at temperatures as high as 1000 K and for durations up to 10 ns.

Model Construction and Physical Prior

Unlike popular Behler‑Parrinello‑type MLPs, the proposed GP model is trained on pre‑computed atomic energies that are directly derived from strict quantum‑mechanical laws, making the training data interpretable. A crucial design choice is the inclusion of a prior mean function m, which provides the model with a physically correct starting point, preventing immediate collapse even when molecules are stretched, heated, or shaken.

Static Performance

Figure 1 demonstrates that the GP atomic energy model attains the expected high accuracy in static tests, with robustness metrics matching theoretical expectations.

Robustness Under Extreme Geometries

The team evaluated the models using highly unstable high‑energy starting geometries (SG). Only the MF5 model (the GP model) successfully relaxed these non‑physical structures and completed 1 ns simulations at four different temperatures. In contrast, MF1 and MIN models collapsed within a few thousand steps. Analysis of the force vectors showed that the GP model’s forces correctly pointed toward “stretching short bonds and compressing long bonds,” providing direct evidence of restorative behavior.

Long‑Time Stability Tests

To further verify long‑term stability, the MF5 model was run in 50 independent 10 ns simulations at 500 K, totaling 0.5 µs of simulation time. Even highly flexible molecules such as aspirin, serine, and glycine remained stable throughout.

Implications for Molecular Dynamics

The dynamic stability of MLPs depends not only on the coverage of training data but also on the model’s extrapolation expectations. Traditional assumptions that unknown regions have low energy are often wrong; far from equilibrium, configurations can have very high energies. If a model underestimates these energies, it creates spurious attractive forces that cause collapse.

The GP model’s prior mean function m determines whether a simulation will “immediately crash” or run for extended periods, while the quantum‑chemical topology prior introduces an inductive bias that suppresses arbitrary energy fluctuations. This approach promises more reliable long‑time explorations for drug design, new‑material discovery, and simulations under extreme conditions.

For further details, see the original paper “Unprecedented robustness of physics‑informed atomic energy models at and beyond room temperature” (Communications Chemistry, 2026) and related news release at https://www.eurekalert.org/news-releases/1121910.