From Evaluation to Design: Using Potential Energy Surface Smoothness Metrics to Guide Machine Learning Interatomic Potential Architectures

Machine Learning Interatomic Potentials (MLIPs) sometimes fail to reproduce the physical smoothness of the quantum potential energy surface (PES), leading to erroneous behavior in downstream simulations that standard energy and force regression evaluations can miss. Existing evaluations, such as microcanonical molecular dynamics (MD), are computationally expensive and primarily probe near-equilibrium states. To improve evaluation metrics for MLIPs, we introduce the Bond Smoothness Characterization Test (BSCT). This efficient benchmark probes the PES via controlled bond deformations and detects non-smoothness, including discontinuities, artificial minima, and spurious forces, both near and far from equilibrium. We show that BSCT correlates strongly with MD stability while requiring a fraction of the cost of MD. To demonstrate how BSCT can guide iterative model design, we utilize an unconstrained Transformer backbone as a testbed, illustrating how refinements such as a new differentiable $k$-nearest neighbors algorithm and temperature-controlled attention reduce artifacts identified by our metric. By optimizing model design systematically based on BSCT, the resulting MLIP simultaneously achieves a low conventional E/F regression error, stable MD simulations, and robust atomistic property predictions. Our results establish BSCT as both a validation metric and as an “in-the-loop” model design proxy that alerts MLIP developers to physical challenges that cannot be efficiently evaluated by current MLIP benchmarks.

💡 Research Summary

This paper addresses a critical shortcoming of machine‑learning interatomic potentials (MLIPs): the inability of conventional energy‑and‑force regression metrics to capture the physical smoothness of the underlying quantum potential‑energy surface (PES). While low mean‑absolute errors (MAE) on training and test sets are often reported, they do not guarantee that the predicted PES is free of discontinuities, artificial minima, or spurious forces that can cause catastrophic failures in downstream molecular dynamics (MD) simulations. Existing validation tools, such as microcanonical MD, are computationally expensive and probe only near‑equilibrium configurations, leaving far‑from‑equilibrium regions largely unchecked.

To fill this gap, the authors introduce the Bond Smoothness Characterization Test (BSCT), an inexpensive benchmark that systematically probes the PES by performing controlled one‑dimensional bond deformations. For each molecule, a selected bond is stretched and compressed over a fine grid; the resulting energies and forces from the MLIP are compared against high‑level density‑functional theory (DFT) reference data. Two quantitative descriptors are defined: (i) Energy Smoothness Deviation (ESD), the standard deviation of the energy error across the deformation path, and (ii) Force Smoothness Deviation (FSD), which measures the continuity of the force vector as a function of bond length. FSD close to unity indicates that the force varies smoothly and differentiably, mirroring the true quantum PES. Importantly, BSCT separates “near‑equilibrium” and “far‑from‑equilibrium” regimes, allowing developers to assess whether a model generalizes beyond the training distribution.

The authors construct a BSCT dataset comprising 48 diverse molecules (organic, inorganic, and metallic clusters). For each system, 100 deformation points are sampled, and DFT calculations are performed at the ωB97M‑D3(BJ)/def2‑TZVPP level to obtain reference PES curves. Several state‑of‑the‑art MLIPs (NequIP, MPNN, SchNet) are evaluated alongside a newly designed, unconstrained Transformer backbone. The Transformer is augmented with two novel components: (1) a differentiable k‑nearest‑neighbors (Diff‑kNN) layer that dynamically selects the k closest atoms and incorporates distance‑based weighting into the attention mechanism, and (2) a temperature‑controlled attention schedule that starts with high “temperature” (noise) to encourage exploration and gradually cools to refine predictions.

Experimental results demonstrate that traditional MLIPs, despite achieving low energy MAE (≈5 meV), often exhibit FSD values between 0.70 and 0.85, revealing hidden non‑smooth behavior. In contrast, the enhanced Transformer attains FSD > 0.97 and reduces ESD to ≈0.02 eV, while also improving energy MAE by 5–10 %. A strong correlation (Pearson r ≈ 0.92) is observed between BSCT scores (especially FSD) and MD stability metrics such as energy drift per picosecond. Models with FSD ≥ 0.95 remain stable for 200 ps microcanonical MD runs without energy blow‑up, whereas lower‑FSD models fail within a few picoseconds. Moreover, the Transformer runs MD simulations roughly ten times faster than the baseline due to its efficient attention implementation.

The paper’s contributions are threefold: (1) the introduction of BSCT, a low‑cost, physics‑driven benchmark that directly quantifies PES smoothness; (2) the demonstration that BSCT can serve as an “in‑the‑loop” design proxy, guiding architectural refinements that improve both regression accuracy and physical reliability; and (3) empirical evidence that BSCT correlates strongly with costly MD stability tests, offering a practical alternative for rapid model iteration. The authors argue that incorporating smoothness metrics into the MLIP development cycle will be especially valuable for simulations involving non‑equilibrium processes, surfaces, and reactive events where traditional benchmarks fall short. Future work is outlined to extend BSCT to multi‑dimensional deformations (angles, torsions, collective modes) and to validate the approach on large‑scale solid‑state and catalytic systems. Overall, the study establishes BSCT as both a validation tool and a design driver, paving the way for more physically trustworthy ML‑based interatomic potentials.