Compact atomic descriptors enable accurate predictions via linear models.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Condensed matter physics

Background:

• Accurate prediction of material properties like formation energies and atomic forces is crucial for materials discovery.
• Current machine learning methods often require complex descriptors and significant computational resources.

Purpose of the Study:

• To evaluate the accuracy of linear ridge regression with a three-body local density representation for predicting material properties.
• To explore methods for descriptor sparsification and computational efficiency improvement.
• To assess the potential for developing compressed, material-agnostic descriptors.

Main Methods:

• Utilized a three-body local density representation derived from the atomic cluster expansion.
• Employed linear ridge regression for fitting formation energies and atomic forces.
• Applied principal component analysis (PCA) and least absolute shrinkage operator (LASSO) regression for descriptor sparsification.

Main Results:

• Linear ridge regression achieved accuracy comparable to state-of-the-art, more complex machine learning methods.
• Descriptor sparsification using PCA and LASSO reduced descriptor size by four times with maintained or improved accuracy.
• Reduced descriptors exhibited shared features across multiple datasets, indicating potential for generalizability.

Conclusions:

• Simple linear ridge regression offers a computationally efficient and accurate alternative for materials property prediction.
• Descriptor compression is a viable strategy to enhance computational efficiency without sacrificing predictive power.
• The findings suggest the feasibility of creating universally applicable, compressed, and accurate material descriptors.