Abstract Machine-learned potentials (MLPs) have become widely adopted alternatives to traditional electronic structure and molecular mechanics methods. However, despite excelling on fixed test sets, most MLPs remain prone to instability when deployed in molecular dynamics simulations, particularly at elevated temperatures. Here, we present the first physics-informed Gaussian process (GP)-based atomic energy models that achieve practically unlimited stability in NVT simulations at temperatures as high as 1000 K. Our findings highlight the importance of the GP prior mean function and demonstrate the models’ ability to predict restoring forces that preserve the system’s physical integrity. The quantum chemical topology information embedded in these models acts as an inductive bias to mitigate arbitrary fluctuations in the predicted atomic energies. Finally, the models’ robustness is evidenced by 50 successful simulations of four flexible organic molecules (peptide-capped glycine and serine, malondialdehyde, and aspirin), yielding a cumulative simulation time of 0.5 microseconds completed within two CPU days.
Isamura et al. (2026) studied this question.