Robustness of Local Predictions in Atomistic Machine Learning Models.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning models for molecules and materials often decompose global properties into local, atom-centered contributions for computational efficiency and interpretability.
• While practical, the rigorous definition and robustness of these local contributions are not guaranteed.
• Sensitivity to training strategies and model architecture necessitates careful consideration.

Purpose of the Study:

• To introduce a quantitative metric, local prediction rigidity (LPR), for assessing the robustness of locally decomposed predictions in machine learning models.
• To investigate how training data composition influences LPR.
• To develop strategies for enhancing LPR to improve model performance.

Main Methods:

• Development and application of the local prediction rigidity (LPR) metric.
• Systematic investigation of LPR's dependence on training data set composition.
• Testing across various problems, from toy models to real chemical systems.

Main Results:

• Demonstrated the utility of LPR in quantifying the robustness of atom-centered contributions in machine learning models.
• Identified key aspects of model training, particularly data set composition, that affect LPR.
• Showcased strategies to systematically improve LPR.

Conclusions:

• Local prediction rigidity (LPR) provides a crucial measure for the reliability of atomistic machine learning models.
• Strategies to enhance LPR can significantly improve the robustness, interpretability, and transferability of these models.
• This work offers a pathway to more trustworthy and interpretable machine learning in chemistry and materials science.