CatTSunami: Accelerating Transition State Energy Calculations with Pretrained Graph Neural Networks
Brook Wander, Muhammed Shuaibi, John R. Kitchin, Zachary W. Ulissi and C. Lawrence Zitnick, “CatTSunami: Accelerating Transition State Energy Calculations with Pretrained Graph Neural Networks”, ACS Catalysis 15(7), 5283-5294 (2025). https://doi.org/10.1021/acscatal.4c04272
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Transition state energies are central to catalytic kinetics, but nudged elastic band (NEB) calculations with DFT are expensive, so only a handful of the many possible adsorbate and site configurations are usually tried, and screening relies instead on Brønsted-Evans-Polanyi (BEP) relations or other approximations. Earlier machine-learning accelerations of NEB needed a bespoke model for each system. The paper asks whether graph neural network potentials pretrained only on the Open Catalyst 2020 (OC20) adsorbate relaxation data, with no reaction or transition state data, can be used off the shelf (zero-shot) to find transition states close to DFT, and what that enables (Wander et al. 2025).
Approach
The authors built CatTSunami, a high-throughput framework (released in the fairchem repository) that places reactants and products on a surface in many configurations, relaxes them with an ML potential, assembles initial and final NEB frames favoring low ML energy and short adsorbate displacement, and interpolates the band with IDPP-like corrections and careful unwrapping across periodic boundaries. To benchmark it they created OC20NEB, 932 RPBE NEB calculations in VASP spanning desorptions, dissociations and transfers (about 300 per class, from hand-curated lists of 19, 54 and 25 reactions) on randomly generated surfaces with up to three elements, half on materials held out of OC20 training. NEBs used ten frames, ASE with BFGS, a 1 eV/Ų spring constant, climbing image switched on later, and a 0.05 eV/Šforce criterion. A result counted as a success if the activation energy was within 0.1 eV of DFT (or both were barrierless). Five pretrained models were tested: EquiformerV2 with 153M and 31M parameters, GemNet-OC, PaiNN and DimeNet++. Four ways of mixing ML and DFT were compared, from all-ML to an ML pre-relaxed band finished with DFT.
Key findings
Success tracked model quality on OC20. With DFT-relaxed end states, an ML NEB and a DFT single point on the transition state (the “ML + 2 DFT RXs + 1 DFT SP” scheme), the 153M EquiformerV2 reached 98.8% success for desorptions, 89.9% for dissociations and 83.8% for transfers, whereas PaiNN and DimeNet++ fell to roughly 12 to 28% for dissociations and transfers. For the 31M model, mean absolute errors were 0.03, 0.06 and 0.09 eV for the three classes. There was no appreciable difference between in-domain and out-of-domain materials. Across schemes with the 31M model, all-ML NEBs gave about a 2200x speedup with 70% average success; ML relaxations plus three DFT single points gave 84% at 88x; the scheme above gave 88% at 28x; and an ML pre-relaxed band refined by DFT gave DFT accuracy at about 3x, and sometimes found transition states more than 0.1 eV lower than pure DFT NEBs (7% of transfers, 2% of dissociations). Integrated RMSE along the reaction coordinate showed that most ML and DFT paths were structurally similar.
In the first case study, the authors revisited CO hydrogenation on Rh(111), a network of 61 intermediates and 174 dissociation or association reactions previously treated with BEP relations and a Gaussian process. CatTSunami found transition states for 162 of the 174 reactions using about 19,000 NEBs in 12 GPU days, estimated at 52 GPU years with DFT (about 1500x). Against 102 converged DFT NEBs, BEP relations fit to those same data gave an MAE of 0.20 eV, while ML plus three DFT single points gave 0.04 eV, and the 0.75 eV feasibility cutoff changed which reactions were deemed accessible. In the second, a microkinetic ammonia synthesis volcano on stepped Ru, Fe, Pd, Co, Ni and Rh was reproduced in about 10 GPU hours (versus an estimated 1.7 GPU years with DFT). Systematically lower energy intermediates and N2 dissociation transition states shifted the volcano and placed Ru more clearly at its apex, in line with experiment. Because OC20 lacks spin polarization, Co and Ni needed extra spin-polarized relaxations and dimer refinements.
This summary is based on the arXiv v3 preprint text (June 2024), including its supplementary tables, rather than the final published version.
Significance
The work shows that a general adsorbate potential learns the potential energy surface well enough to locate saddle points it was never trained on, making exhaustive enumeration of transition state configurations affordable. That opens the way to replacing BEP estimates with explicit barriers in reaction network studies and to screening catalysts on kinetics directly. It continues the group’s line of ML-accelerated NEB from system-specific surrogates (Yang et al. 2021) toward pretrained foundation potentials, and connects to later use of such potentials for Hessians and vibrational analysis (Wander et al. 2025).