Research map / Probabilistic and scientific modeling / Molecules and proteins

Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 33 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes testing MLIPs by energy conservation in NVE molecular dynamics and introduces eSEN, a smooth, conservative equivariant interatomic potential that achieves state-of-the-art Matbench-Discovery, thermal conductivity, and MDR phonon results.

Problem

Lower test-set errors for machine learning interatomic potentials (MLIPs) do not reliably translate into accuracy on downstream physical property prediction tasks such as relaxation, MD, and phonon/thermal conductivity calculations.

Approach

Use energy conservation in NVE MD simulations (with bounded, smooth energy derivatives) as a practical test of whether an MLIP is suitable for property prediction; design eSEN with conservative forces, no discretized nonlinearities (gated SiLU in spherical-harmonic space), no neighbor limit, and envelope functions so that test error correlates better with property accuracy.

Key results

• On compliant Matbench-Discovery, eSEN-30M-MP achieves F1 0.831 and kappa_SRME 0.340, the best among MPtrj-only models.
• On non-compliant Matbench-Discovery, eSEN-30M-OAM achieves F1 0.925 and kappa_SRME 0.170 with RMSD 0.0608.
• On SPICE-MACE-OFF, eSEN-6.5M attains lower energy and force MAE than MACE-OFF-L (4.7M) and EScAIP-45M on all splits.

Measured effects reported in the paper

Surprising findings

• Direct-force models can achieve SOTA test accuracy while causing significant energy drift in MD and large phonon errors at small displacements.
• Direct-force models showed decreasing phonon MAE with larger finite displacements (0.2 Å), suggesting the benchmark result depends on displacement choice.

Limitations

• Direct-force pre-training followed by conservative fine-tuning is needed to reach top accuracy; direct-force models are not conservative.
• Ablation experiments used small 2-layer 3.2M models, and phonon comparisons to OMat-trained models were not made due to level-of-theory mismatch.
• Energy conservation is a proxy; the correlation between test error and property accuracy holds only among models that pass the conservation test.

Experimental setup

Datasets: MPTrj SPICE-MACE-OFF OMat24 sAlex TM23 MD22 MDR Phonon benchmark SPICE

Benchmarks: Matbench-Discovery MDR Phonon benchmark SPICE-MACE-OFF test set TM23 energy conservation MD22 energy conservation

Metrics: F1 kappa_SRME RMSD MAE energy drift DAF

Baselines: EquiformerV2 / eqV2-S-DeNS MACE-MP-0 CHGNet M3GNet SevenNet-l3i5 GRACE-2L ORB v2 MACE-OFF-L EScAIP-45M eqV2-M-OAM

Builds on: eSCN Equiformer DimeNet DeNS MPTrj

Closest papers by meaning

• MLIP Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform (NeurIPS 2025 · similarity 0.90)
• Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.89)
• Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials (NeurIPS 2026 · similarity 0.89)
• Learning the Electronic Hamiltonian of Large Atomic Structures (ICML 2025 · similarity 0.86)
• The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.86)
• The dark side of the forces: assessing non-conservative force models for atomistic machine learning (ICML 2025 · similarity 0.86)