Prediction rigidities for data-driven chemistry.

Area of Science:

• * Computational Chemistry
• * Machine Learning in Chemical Sciences

Background:

• * Machine learning (ML) is increasingly vital in chemical sciences for correlating structures with properties.
• * Understanding ML model learning, interpretability, and transferability is crucial for efficient application.
• * Current methods for assessing ML model behavior lack detailed local insights.

Purpose of the Study:

• * To introduce and demonstrate the utility of prediction rigidities for analyzing ML models in chemistry.
• * To assess ML model robustness at both global and local prediction levels.
• * To guide dataset construction and improve ML model training efficiency and interpretability.

Main Methods:

• * Derivation of prediction rigidities as a family of metrics from the ML model's loss function.
• * Application of prediction rigidities to evaluate ML model performance and learning behavior.
• * Implementation of prediction rigidities for coarse-grained ML models in atomistic simulations.

Main Results:

• * Prediction rigidities effectively assess ML model robustness across global and component-wise prediction levels.
• * These metrics provide insights into the learning dynamics of various ML models.
• * The study demonstrates improved dataset construction strategies guided by prediction rigidity analysis.
• * Applicability of prediction rigidities is shown for coarse-grained ML models.

Conclusions:

• * Prediction rigidities offer a powerful tool for understanding and improving ML models in chemical applications.
• * The metrics enhance interpretability and transferability of ML models by revealing local prediction behaviors.
• * This approach facilitates more efficient dataset curation and model development for chemical structure-property prediction.