Designing hybrid descriptors for improved machine learning models in atomistic materials science simulations

The efficiency and performance of machine learning force fields in materials science are significantly influenced by the quality of the embedding of the atomic neighborhood through atomic descriptors and learning algorithms. In this study, we explore various methods for combining two existing descriptors to formulate a hybrid descriptor that enhances the performance of the learned force field. Within the framework of widely used models in materials science, such as linear and quadratic models, we propose several strategies for the design of hybrid descriptors and we perform an extensive analysis to evaluate their performance. Using databases for defects in body-centered cubic (W, Fe) and face-centered cubic (Ni) metals, we assess the models’ effectiveness. Additionally, we explore their potential as surrogate models for challenging yet informative quantities in atomistic materials science, such as the vibrational entropy of defects. By improving the coupling scheme of both descriptors, we propose models that combine the advantages of linear models, which are linear in descriptor dimensionality and easy to train, with the precision of quadratic models, which have quadratic dimensionality and are more challenging to parameterize. The proposed approaches have numerical complexity similar to linear models while maintaining the accuracy of quadratic models, opening new possibilities for machine learning-based molecular dynamics and surrogate modeling in materials science.