(248e) Evaluating Machine-Learned Interatomic Models through Graph-Based Descriptor Analysis

Atomic-scale simulations are vital for advancing chemical and materials research, providing intricate, experimentally inaccessible mechanistic and kinetic insights. However, their effectiveness hinges on underlying interatomic models (IAMs) used to describe the system’s potential energy surface. Quantum-based methods offer high accuracy but are computationally demanding for systems larger than a few hundred atoms whereas classical “force field” methods are orders of magnitude more computationally efficient but have limited accuracy. Machine-learned interatomic models (ML-IAM) provide a powerful means of bridging this accuracy and efficiency gap by learning suitable IAM from sufficient training data. However, the development of IAMs face hurdles including the need for large informative training sets, improved automation, and uncertainty quantification. A lack of suitable metrics for quantitative description and comparison of simulation configurations are critical for addressing these challenges. In this presentation, we describe efforts to develop innovative physics informed graph-based “fingerprinting” methods to fill this capability gap. Notably, our approach provides insight into how ML-IAMs perceive system structure and lends new physical insights into particularly complex chemical and material systems.