Title: Leveraging Potential Energy Surface Smoothness to Shape Machine Learning Interatomic Potential Architectures

Abstract:

Machine Learning Interatomic Potentials (MLIPs) often struggle to replicate the physical smoothness inherent in quantum potential energy surfaces (PES). This deficiency can cause inaccurate outcomes in downstream simulations—errors that traditional evaluations based on energy and force regression frequently overlook. Current assessment methods, such as microcanonical molecular dynamics (MD), are not only computationally intensive but also focus predominantly on states near equilibrium.

To address these limitations, we propose the Bond Smoothness Characterization Test (BSCT), a more efficient benchmark designed to enhance MLIP evaluation. By subjecting the PES to controlled bond deformations, BSCT identifies non-smooth features, including discontinuities, artificial minima, and spurious forces, across both equilibrium and non-equilibrium regions. Our analysis reveals that BSCT scores are strongly correlated with MD stability, yet they require significantly less computational resources than full MD simulations.

We further demonstrate the utility of BSCT as a guide for iterative model development. Using an unconstrained Transformer backbone as a test case, we show that incorporating specific refinements—such as a novel differentiable $k$-nearest neighbors algorithm and temperature-controlled attention mechanisms—effectively mitigates artifacts flagged by our metric. Through systematic model optimization driven by BSCT, the resulting MLIP achieves low conventional error rates for energy and force regression, ensures stable MD simulations, and delivers robust predictions of atomistic properties. These findings position BSCT as a dual-purpose tool: a validation metric for practitioners to gauge MLIP reliability and an "in-the-loop" design proxy that highlights physical complexities overlooked by existing benchmarks. The BSCT dataset and evaluation code are publicly accessible at https://github.com/ryanliu30/bsct.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC