Toward Controllable Catalyst Inverse Design via Large-Scale Autoregressive Pretraining

In three linesGenerative model based on GPT architecture for inverse design of heterogeneous catalysts. Pretrained on 133 million structures, fine-tuned on ~460,000 optimized structures. Achieves 98% structural validity, 95% optimization validity, and improves screening efficiency 1.5–4× for reaction-targeted catalyst discovery.

Summary generated by Claude — human-verified