AETHER

The problem

Finding catalysts for the hydrogen evolution reaction through DFT calculations is expensive and slow. The practical question is whether a model pre-trained on millions of structures can replace that calculation within an acceptable error, and whether the advantage holds when the experiment is repeated across different seeds.

My contribution

I built a reproducible pipeline for curation, training, comparison and explainability over 5,860 Catalysis Hub structures, evaluated on two splits, a random one and one that keeps bulk compositions apart between training and test, to measure how well models generalize to materials they have not seen. I compared four approaches on the same data (physicochemical descriptors with Extra Trees, SchNet trained from scratch, an MLP head on the frozen pretrained MACE-MP-0, and Extra Trees on MACE embeddings), reporting mean and standard deviation over 5 seeds. SHAP-based descriptor selection is done without leaking the test set, and the MAE difference between models comes with a bootstrap confidence interval. The best result, an ensemble of 5 SchNets, reaches an R² of 0.9734 and an MAE of 0.0585 eV. The models are served by a FastAPI API with a Vue 3 interface that estimates a structure's ΔG_H, and the software's INPI registration is in progress.

Project highlights

• 5,860curated structures
• 0.9734best ensemble R²
• 5 seedsreproducible evaluation

Technologies

Python, PyTorch, PyTorch Geometric, MACE, SchNet, SHAP, FastAPI