Towards Material Design with AI Foundation Models—Ligand Selection for Better LED Materials and Gas Absorption Prediction

Recent advancements in AI have popularized the "pre-training/fine-tuning" paradigm in large model development, driving significant progress across various domains. Specifically, Uni-Mol, developed by DP Technology, is a powerful and novel framework that has already shown its capability in 3D molecular conformation and energy property prediction. In two recent works, we expanded the application of Uni-Mol beyond the small molecules to a broader area of material design. As our initial attempt, we constructed a hierarchical high throughput virtual screening (HTVS) workflow including a structure generator, structure filter, Uni-Mol property predictor, and Quantum Mechanics calculator to screen over millions of organic ligand candidates for the Ir(III)-based LED material. Uni-Mol shows promising performance in predicting various electronic and photonic properties, and our HTVS workflow shows its potential by covering a wide range of spectrum to provide candidates for different usages. In our second attempt, Uni-Mol was formulated to robustly depict the Metal-Organic Framework (MOF) and Covalent Organic Framework (COF) structures, and extra input feature blocks were added to host the information from guest molecules (i.e., gas absorbates that interact with the framework) and environmental conditions (i.e., temperature and pressure). This model accurately predicts the gas absorption for systems in various databases and can also make satisfying extrapolate predictions to account for unknown absorbents and pressure conditions.