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

Machine Learning Interatomic Potentials (MLIPs) have successfully demonstrated power-low scaling in their 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 model by itself has acquired the knowledge of bond dissociation energy (BDE) for archetypal bond types. The emergent BDE values quantitatively agree with literature and are found to be robust across distinct organic and inorganic training sets. E3D employs a set of internal representations, probability distribution, and associated information entropy to enable visual inspection and quantitative assessment of various model training scenarios. We apply E3D framework and discuss the synergetic effect of hybrid training set along with its potential to overcome the scaling wall for transition state energy prediction problem.