Fine-tuning universal machine learning potentials for transition state search in surface catalysis
Raffaele Cheula, Mie Andersen and John R. Kitchin, “Fine-tuning universal machine learning potentials for transition state search in surface catalysis”, npj Computational Materials (2026). https://doi.org/10.1038/s41524-026-02228-1
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.
Note: the extracted text is the unedited “Article in Press” version, so volume and page numbers are not yet assigned. Supplementary figures (S1 to S9) are referenced but not included in the text.
Question
Locating transition states (first-order saddle points) of surface reactions with DFT is expensive, task-specific machine learning potentials transfer poorly to new metals or reactions, and universal potentials pretrained on near-equilibrium data are not accurate enough for reactive, bond-breaking configurations. The paper asks whether a robust single-ended saddle search combined with active-learning fine-tuning of a universal potential can deliver DFT-quality transition states at a small fraction of the DFT cost (Cheula et al. 2026).
Approach
The test set is 250 transition states from the reverse water-gas shift network (CO2 hydrogenation to CO and H2O) on metal and single-atom alloy surfaces, previously computed with DFT NEB by Cheula and Andersen. First, using cheap universal potentials without fine-tuning, the authors benchmark four single-ended methods in ASE: Dimer, ARPESS (ClimbFixInternals), Sella, and a new variant they call Bonds-Aware Sella (BA-Sella). In BA-Sella a bond-direction vector is built by summing unit vectors along the adsorbate bonds that form or break between initial and final state (surface bonds excluded, signs chosen so forming and breaking contributions add). When the lowest Hessian eigenmode overlaps this vector by less than 0.5 (absolute dot product), the approximate Hessian is lowered along the bond direction by repeated rank-one updates until the overlap reaches the threshold. Initial guesses come from the highest of eight IDPP-interpolated images. A stochastic restart scheme perturbs the guess (rigid translation with 0.25 Å and per-atom displacements with 0.15 Å standard deviations) and Sobol-sampled optimizer settings. Five universal potentials are compared: CHGNet-0.3, MACE-MPA, eSCN-OC20, eSEN-OAM and UMA-M.
Second, eSEN-OAM is fine-tuned inside an active-learning loop: MLP-driven BA-Sella search, a DFT single point on the resulting structure, fine-tuning on forces only (100 epochs, learning rate 10-4, starting each time from the fresh pretrained model), repeated until DFT forces drop below threshold. A sequential strategy specializes one model per transition state; a batch strategy pools DFT data across all searches into one shared model. These are compared with full DFT optimization and DFT optimization after MLP pre-optimization. DFT used Quantum ESPRESSO with BEEF-vdW and SSSP pseudopotentials.
Key findings
With eSEN-OAM, BA-Sella reached the highest success rate (88%) for finding the intended transition state, against about 80% for Sella, 74% for Dimer and 68% for ARPESS, and Sella-based methods needed far fewer single-point evaluations. The ranking held for all five potentials: the three FAIRChem models gave BA-Sella success of 87 to 90% and CHGNet-0.3 and MACE-MPA about 84%. Pre-refining guesses with NEB did not help much (only about 70% of NEBs converged even after 500 steps), indicating guess quality is not the bottleneck, while up to 20 stochastic restarts raised BA-Sella success to about 97%.
In the DFT-cost comparison, full DFT optimization averaged about 102 DFT single points per structure with a long tail to several hundred, MLP pre-optimization then DFT about 70, batch active learning about 38, and sequential active learning about 8, tightly distributed. Sequential active learning matched the reference NEB transition state (within 0.1 eV) in a fraction similar to or slightly above full DFT optimization; batch was slightly less consistent. The original NEB dataset required roughly 250 steps per NEB, about 2000 DFT single points per transition state, so the sequential workflow is two to three orders of magnitude cheaper. Simply evaluating MLP-found transition states with one DFT single point matched the reference in about 70% of cases. The authors attribute the efficiency of the sequential approach to deliberate overfitting to one local region of the energy surface, which sacrifices transferability; the batch model is more reusable. BA-Sella also improved success over Sella and Dimer for gas-phase reactions and zeolite hydrocarbon chemistry, and slightly for cluster reconstructions when reformulated with all bond-length changes.
Significance
The work shows that universal potentials, made task-accurate by a few targeted DFT single points, can make DFT-quality saddle point searches routine, opening high-throughput barrier calculations across large reaction networks, surfaces and sites for microkinetic modeling. The authors caution that BA-Sella relies on knowing the reaction coordinate in advance and may be less effective for multistep, solvent-mediated or collective lattice rearrangements. Code (arkimede, mlpsfinetuning) and data are released on GitHub.