szl666/t2mat

Text-to-Material

T2MAT(Text-to-Material)

A comprehensive framework processing from a user-input sentence, generating material structures aligning with goal properties beyond the existing database via globally exploring chemical space, followed by an entirely automated Density Functional Theory (DFT) validation workflow.

In the current version of T2MAT, we developed a modified CDVAE with improved symmetry for generating material structures. The CGTNet is a new GNN with improved accuracy and small data friendliness compared to CGCNN, SchNet, DimeNet++, PaiNN and GemNet. The complete version of T2MAT including local LLM, CGTNet, modified CDVAE and automatic DFT framework will be updated after the manuscript of T2MAT is accepted.

Furthermore, we have developed a new generative model——SymGEN. The symmetry(ratio of structures with non-P1 space group, Figure e), thermodynamic stability(DFT-calculated formation energies, Figure f) and kinetic stability(DFT-calculated phonon properties, Figure g) of generated structures are significantly improved compared to the CDVAE.

SymGEN will be replaced with the CDVAE in T2MAT after the manuscript of SymGEN is accepted.