Feature optimization for atomistic machine learning yields a data-driven construction of the periodic table of the...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of atomic-scale properties is crucial for materials discovery and understanding chemical behavior.
• Current machine learning models rely on effective structure representations to capture correlations between composition, structure, and target properties.
• Optimizing these representations can enhance model accuracy and provide chemical insights.

Purpose of the Study:

• To generalize the Smooth Overlap of Atomic Positions (SOAP) kernel for improved machine learning of atomic-scale properties.
• To incorporate distance-dependent weighting and chemical species correlations into the SOAP kernel.
• To enhance the performance of machine learning models for molecular and materials stability.

Main Methods:

• Generalization of the SOAP kernel by introducing a distance-dependent weight.
• Inclusion of a description for correlations between chemical species within the kernel.
• Application of the enhanced kernel to machine learning models for molecular and materials stability.

Main Results:

• Substantial performance improvement in machine learning models for molecular and materials stability.
• Facilitation of analysis for complex, multi-component systems.
• Successful extension of the SOAP kernel to coarse-grained intermolecular potentials.
• Rediscovery of periodic table-like element correlations through optimized representations.

Conclusions:

• The generalized SOAP kernel significantly enhances machine learning accuracy for atomic-scale properties.
• The method provides a powerful tool for analyzing complex chemical systems and developing new materials.
• Machine learning can rediscover and generalize fundamental chemical concepts, bridging data-driven insights with established chemical intuition.