Benchmarking Machine Learning Descriptors for Crystals

The success of machine learning (ML) in materials design and innovation largely hinges on the quality and comprehensiveness of the representation of atoms, molecules, and materials as features. When these features are represented in numerical or vector form, they are known as descriptors. The quality of these descriptors is assessed by their ability to comprehensively capture the physics of chemical and materials systems. Crystal systems are at the heart of materials science, and their periodic and complex structure poses a unique challenge for feature representation. In this study, we benchmark descriptors from the matminer library, the smooth overlap of atomic positions (SOAP) descriptors as implemented in Schrödinger’s Materials Science Suite (MSS), and crystal graph convolutional neural networks (CGCNN) for prediction of three different materials properties. These include the bulk modulus of semiconductors, heat of formation of perovskites, and CO2 adsorption in metal-organic frameworks (MOFs). In the process, we evaluate and compare the performance of these descriptors in terms of the predictive performance of the ML algorithm, ease of use, time, memory, and data intensiveness. In addition, we illuminate the strengths and weaknesses of each of these descriptors along with their cost-benefit trade-off. This benchmarking study gives insights into what descriptors to use for different types and sizes of crystals and provides end-to-end examples of ML pipelines for crystal systems. This is a good starting point for further exploratory ML studies, especially for MOFs, which have environmental benefits and are hitherto less explored.