J Am Chem Soc. 2024 Sep 12. doi: 10.1021/jacs.4c05159. Online ahead of print.

ABSTRACT

There is a growing need to develop artificial intelligence technologies capable of accurately predicting the properties of materials. This necessitates the expansion of material databases beyond the scope of density functional theory, and also the development of deep learning (DL) models that can be effectively trained with a limited amount of high-fidelity data. We developed a DL model utilizing a crystal structure representation based on the orbital field matrix (OFM), which was modified to incorporate information on elemental properties and valence electron configurations. This model, effectively capturing the interrelation between the elemental properties in the crystal, was coined the PRoperty-networking Orbital Field maTrix-convolutional neural Network (PROFiT-Net). Remarkably, PROFiT-Net demonstrated high accuracy in predicting the dielectric constant, experimental band gaps, and formation enthalpies compared with other leading DL models. Moreover, our model accurately identifies physical patterns, such as avoiding the prediction of unphysical negative band gaps and exhibiting a Penn-model-like trend while maintaining the scalability. We envision that PROFiT-Net will accelerate the development of functional materials.

PMID:39264687 | DOI:10.1021/jacs.4c05159