Research map / Probabilistic and scientific modeling / Molecules and proteins

Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 0 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.

Introduces GMD, a benchmark of four compositional-generalisation tasks on small H/C/N/O molecules (PBE labels) for machine learning interatomic potentials, showing OOD errors often far exceed ID errors for scratch-trained and fine-tuned foundation MLIPs.

Problem

Whether machine learning interatomic potentials (MLIPs), including fine-tuned foundation models, generalise compositionally to unseen molecules rather than merely interpolating training examples.

Approach

A controlled benchmark of alkyl-chain and functional-group molecule splits (chain extension, fragment fusion, duplication, combination) where training molecules are chosen so that compositional generalisation to OOD test molecules should be feasible, plus an open toolkit for generating AIMD trajectories.

Key results

• On Fragment Fusion, MACE trained from scratch has OOD force error more than an order of magnitude above its ID error, and OOD energy error more than two orders of magnitude above ID.
• On Fragment Duplication, MACE from scratch has OOD energy error about three orders of magnitude worse than ID.
• Fine-tuned foundation models (MACE-MP-0, UMA-Small) do not consistently outperform MACE trained from scratch; UMA-Small matches ID and OOD on Fragment Combination.

Surprising findings

• Fine-tuned foundation models do not consistently outperform models trained from scratch.
• In-distribution accuracy does not reliably predict out-of-distribution performance.

Limitations

• Scoped to small H/C/N/O molecules in vacuum with PBE labels; may not transfer to aromatic, charged, inorganic, or transition-metal chemistry.
• Foundation models evaluated only under full fine-tuning at a fixed 2000-epoch budget.
• Failure mechanism within architectures not localised; parameter-efficient and readout-only fine-tuning not compared.

Experimental setup

Datasets: GMD (constructed alkyl and functional-group molecules, PBE/def2-TZVP+D3BJ labels)

Benchmarks: GMD Fragment Chain Extension GMD Fragment Fusion GMD Fragment Duplication GMD Fragment Combination

Metrics: force MAE energy MAE cosine similarity force magnitude MAE per-atom energy MAE

Baselines: SchNet GemNet DimeNet++ PaiNN NequIP eSCN EquiformerV2 MACE UMA-Small MACE-MP-0

Builds on: MACE MACE-MP-0 NequIP ASE

Closest papers by meaning

• Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.89)
• The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.89)
• Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction (ICML 2025 · similarity 0.89)
• MLIP Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform (NeurIPS 2025 · similarity 0.88)
• Higher-Rank Irreducible Cartesian Tensors for Equivariant Message Passing (NeurIPS 2024 · similarity 0.87)
• The dark side of the forces: assessing non-conservative force models for atomistic machine learning (ICML 2025 · similarity 0.87)