Machine learning with force-field inspired descriptors for materials: fast screening and mapping energy landscape

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Current machine learning (ML) models for materials science often rely on chemical-only descriptors, limiting their ability to distinguish between different structural prototypes.
• Accurate prediction of material properties is crucial for efficient material screening and understanding energy landscapes in multicomponent systems.

Purpose of the Study:

• To introduce a comprehensive set of chemo-structural descriptors that extend the applicability of ML in materials screening.
• To enable differentiation between structural prototypes, improving ML model accuracy for diverse material systems.
• To facilitate the discovery of new materials with desired properties and validate ML models against established computational methods.

Main Methods:

• Development and application of novel chemo-structural descriptors including pairwise radial, nearest neighbor, bond-angle, dihedral-angle, and core-charge distributions.
• Training of ML models using the gradient boosting algorithm on a dataset of 24,549 bulk and 616 monolayer materials from the JARVIS-DFT database.
• Integration of the formation energy ML model with a genetic algorithm for structure search and validation against DFT convex hull calculations.

Main Results:

• Demonstrated the importance of the combined chemo-structural descriptors in accurately predicting formation energies, bandgaps, refractive indices, magnetic properties, and elastic modulus for 3D materials.
• Achieved highly accurate ML models for predicting properties of both 3D and 2D layered materials, including exfoliation energies.
• Successfully utilized trained ML models to discover exfoliable 2D materials meeting specific property criteria.

Conclusions:

• The proposed chemo-structural descriptors significantly enhance the predictive power of ML models for a wide range of material properties.
• The developed ML models provide accurate predictions and enable efficient screening and discovery of novel materials, particularly 2D layered systems.
• The integration with genetic algorithms and validation against DFT convex hull offers a robust framework for evaluating and applying ML models in materials science.