Metric for physics awareness of machine-learning potentials

Our manuscript arXiv:2408.00755 presents a theoretical and computational framework to quantify the physics-awareness of machine-learning potentials (MLP), connecting interatomic forces to physical observables such as the thermal conductivity and other thermomechanical properties. This work shows that established metrics based on formation energy can be inaccurate in predicting crystal structures, interatomic forces, and thermal conductivity. Because it tests the second- and third-order derivatives of the potential energy surface (PES), and higher derivatives expose subtle discontinuities in the PES, this benchmark is becoming the standard to measure how the physical accuracy of the PES encoded by a MLP affects technologically relevant macroscopic observables.

Our work has received recognition on the matbench discovery platform [https://matbench-discovery.materialsproject.org], and has also been employed by companies such as Meta [http://arxiv.org/abs/2502.12147], Preferred Networks [https://tech.preferred.jp/en/blog/lattice-thermal-conductivity-calculation-with-pfp/], and Orbital Materials [https://arxiv.org/abs/2504.06231].