New InterCat paper - Raffaele Cheula and Mie Andersen
Title: Fine-tuning universal machine learning potentials for transition state search in surface catalysis
Can universal machine learning potentials make transition-state searches fast enough for high-throughput catalyst screening, without sacrificing DFT accuracy?
This is the question we address in our new paper, now published in npj Computational Materials
Finding transition states is essential for predicting reaction rates and understanding catalytic mechanisms, but conventional DFT-based searches are computationally expensive. Universal machine learning potentials offer a much faster alternative, but their accuracy can deteriorate in reactive configurations.
Using a dataset of 250 transition states from CO₂ hydrogenation reactions on metal and single-atom alloy surfaces, we combine universal machine learning potentials with transition-state search algorithms and active learning to efficiently reach DFT-quality transition states at a fraction of the computational cost.
Some highlights:
- We introduce Bond-Aware Sella, a modification that substantially improves the success rate of transition-state searches.
- Active learning allows the machine learning potential to be fine-tuned specifically where additional accuracy is needed.
- DFT-quality transition-state structures can be obtained using only ~10 DFT single-point calculations on average per transition state, opening the way toward much more efficient reaction-network and catalyst screening.
The results show how universal machine learning potentials can move beyond zero-shot predictions and become practical tools for accurate, large-scale exploration of catalytic reaction mechanisms.