Beyond Scaling: Chemical Intuition as Emergent Ability of Universal Machine Learning Interatomic Potentials

TL;DR

This paper introduces the E3D framework to study how machine learning interatomic potentials develop the ability to understand chemical bonds, revealing emergent representations of bond dissociation energy that align with known chemistry and are robust across datasets.

Contribution

The paper presents a novel E3D framework and demonstrates that MLIPs can spontaneously learn chemically meaningful representations without explicit supervision, advancing interpretability in molecular modeling.

Findings

MLIPs learn bond dissociation energies aligning with literature

Representations are robust across diverse datasets

Energy decomposability correlates with chemical reactivity

Abstract

Machine Learning Interatomic Potentials (MLIPs) have successfully demonstrated scaling behavior, i.e. the power-law improvement in training performance, however the emergence of novel capabilities at scale remains unexplored. We have developed Edge-wise Emergent Decomposition (E3D) framework to investigate how an MLIP develops the ability to derive physically meaningful local representations of chemical bonds without explicit supervision. Employing an E(3)-equivariant network (Allegro) trained on molecular data (SPICE~2), we found that the trained MLIP spontaneously learned representations of bond dissociation energy (BDE) by decomposing the global potential energy landscape. The learned BDE values quantitatively agree with literature and its scalability are found to be robust across diverse training datasets, suggesting the presence of underlying representation that captures chemical reactions faithfully beyond given training information. Our E3D analysis utilizing Shannon's entropy reveals a close interplay between the decomposability of potential energy learning, scalability of learning, and emergent chemical reactivity, thus providing novel insights of scaling limitations and pathways toward more physically interpretable and predictive simulations.