This project proposes a novel, fully automated dual-loop computational workflow for inverse materials discovery that couples a generative AI model for crystal structure creation with a mixed machine-learned interatomic potential (MLIP) / ab initio (DFT) high-throughput active-learning labeling loop.
The workflow is specifically targeted to discover materials with low surface energy—of high relevance to heterogeneous catalysis, protective coatings and tribo-chemical applications—by using surface energy as a rapid, physics-based descriptor to steer generation and selection. In the outer loop a property-conditioned generative model proposes candidate structures; in the inner loop an ensemble of MLIP models provides fast property estimates and uncertainty metrics that trigger selective DFT labeling only when necessary, enabling efficient exploration of new chemical space.
The project will deliver an open-source end-to-end pipeline, a public database of ~10^5 screened structures with MLIP/DFT labels, and a set of validated, most promising, low-surface-energy candidates for experimental follow-up. The hybrid approach leverages recent advances in generative modeling, accurate MLIPs and active learning to dramatically reduce computational cost while preserving ab initio fidelity.
The workflow is designed for large, contiguous GPU allocations on HPC infrastructure to support distributed model training and the independent DFT evaluations required for robust discovery.