Prediction rigidities for data-driven chemistry

• DOI: 10.1039/d4fd00101j
• Publication Date: 2024-08-23T14:21:26Z
• Authors: Sanggyu Chong; Filippo Bigi; Federico Grasselli; Philip Loche; Matthias Kellner; Michele Ceriotti
• Keywords: Interpretability; Transferability; Leverage (statistics); Robustness; Training set; Formalism (music); Experimental data

The widespread application of machine learning (ML) to the chemical sciences is making it very important understand how ML models learn correlate structures with their properties, and what can be done improve training efficiency whilst guaranteeing interpretability transferability. In this work, we demonstrate wide utility prediction rigidities, a family metrics derived from loss function, in understanding robustness model predictions. We show that rigidities allow assessment not only at global level, but also on local or component-wise level which intermediate (e.g. atomic, body-ordered, range-separated) predictions are made. leverage these behavior different models, guide efficient dataset construction for training. finally implement formalism targeting coarse-grained system applicability an even broader class atomistic modeling problems.