PROFiT-Net: Property-Networking\nDeep Learning Model\nfor Materials

There is a growing need to develop\nartificial intelligence\ntechnologies\ncapable of accurately predicting the properties of materials. This\nnecessitates the expansion of material databases beyond the scope\nof density functional theory, and also the development of deep learning\n(DL) models that can be effectively trained with a limited amount\nof high-fidelity data. We developed a DL model utilizing a crystal\nstructure representation based on the orbital field matrix (OFM),\nwhich was modified to incorporate information on elemental properties\nand valence electron configurations. This model, effectively capturing\nthe interrelation between the elemental properties in the crystal,\nwas coined the PRoperty-networking Orbital Field maTrix-convolutional\nneural Network (PROFiT-Net). Remarkably, PROFiT-Net demonstrated high\naccuracy in predicting the dielectric constant, experimental band\ngaps, and formation enthalpies compared with other leading DL models.\nMoreover, our model accurately identifies physical patterns, such\nas avoiding the prediction of unphysical negative band gaps and exhibiting\na Penn-model-like trend while maintaining the scalability. We envision\nthat PROFiT-Net will accelerate the development of functional materials.