“Few-Shot” Catalyst Design with Physics-Informed Foundation Models

Deng (2023) and Xu et al. (2025) discuss the growing role of large ML models and few-shot learning in catalyst prediction. However, most models require extensive datasets for each new reaction type. My idea: train a “foundation model” on massive, heterogeneous catalyst data (including text, structure, and computed properties), then fine-tune it for new reactions using a handful of computational or experimental examples, incorporating physics-based constraints (e.g., conservation laws, scaling relations from Zhang et al., 2025). This approach would enable fast, reliable catalyst design in data-scarce scenarios, democratizing advanced computational discovery for virtually any chemist. It’s a conceptual transfer from breakthroughs in AI for language and vision to the unique challenges of catalysis.

References:

1. Catalysis distillation neural network for the few shot open catalyst challenge. B. Deng (2023). arXiv.org.
2. Advances in computational approaches for bridging theory and experiments in electrocatalyst design.. Yaqin Zhang, Yu Xiong, Yuhang Wang, Qianqian Wang, Jun Fan (2025). Nanoscale Horizons.
3. AI-Empowered Catalyst Discovery: A Survey from Classical Machine Learning Approaches to Large Language Models. Yuanyuan Xu, Hanchen Wang, Wenjie Zhang, Lexing Xie, Yin Chen, Flora Salim, Ying Zhang, Justin Gooding, Toby Walsh (2025). arXiv.org.