An Investigation into the Preparation of Crystal and Atomic Data for Deep Learning Models

Computationally efficient methods for materials investigation, whether from first principles or deep learning models, have aided and accelerated the process of discovering and designing new materials. In this research, we focus on a deep-learning-based approach, made possible by open and easily searchable materials databases like the Materials Project [1, 2]. Developing an accurate and useful model requires careful selection and preprocessing of input data. In this presentation, we discuss procedures for obtaining and preparing physical, chemical, and structural information for use in predicting the stability and formation energies of intermetallic compounds with the Heusler structure, specifically with the space group 225. We extracted structural information like fractional coordinates and used the information in conjunction with tables of atomic properties of individual elements that could comprise these materials. We collected and sorted the properties of materials with the Heusler structure from select databases of experimental and computational data, augmented this with the properties of constituent elements, to be used as input data for deep learning models.

[1] Jain, A., Ong, S. P., Hautier, G. et al. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater. 1, 011002 (2013).

[2] Horton, M.K., Huck, P., Yang, R.X. et al. Accelerated data-driven materials science with the Materials Project. Nat. Mater. 24, 1522–1532 (2025).