Transition-state search stops assuming the ground state: global sampling plus a 30,000-parameter network remaps propane dehydrogenation
Synopsis
Transition-state searches have long started from hand-chosen ground-state structures, which can miss metastable configurations that may govern rates at finite temperature. A workflow pairs global configurational search with a lightweight graph neural network: it samples many reactant, product and intermediate configurations on the potential energy surface, then predicts transition-state geometries with an edge-aware attention network of about 30,000 trainable parameters. On propane dehydrogenation over Pt4 supported on carbon nitride, the workflow recovers the known mechanism and yields competing pathways absent from earlier hand-built models; the authors also note that the metastable tetrahedral Pt4, though intrinsically more reactive, is too costly to form to dominate actual kinetics.
Interpretation
Transition-state search can now start from a whole configuration space rather than only the lowest-energy ground state, bringing metastable configurations and adsorbate migration steps into one reaction-path map. Earlier transition-state searches typically connected two local energy minima fixed by gradient-based optimization, which under realistic catalytic conditions can miss pathways through metastable configurations populated by finite-temperature fluctuations and linked by lower barriers. For propane dehydrogenation on carbon-nitride-supported Pt4, the global search placed the strongest propane binding at the Pt4-support interface, a configuration absent from earlier studies that built adsorption geometries by hand.
The transition-state predictor is deliberately kept small, with a parameter count matched to the available surface transition-state data rather than trading capacity for accuracy. Existing transition-state prediction models typically carry more than one million trainable parameters, whereas heterogeneous catalysis datasets often hold only about a thousand examples, a mismatch that can fail on chemically distinct systems. The network has 30,089 trainable parameters, pretrained on Transition1x and then fine-tuned on OC20NEB; after fine-tuning, RMSD on OC20NEB fell from 0.445 to 0.235 while RMSD on Transition1x moved only from 0.216 to 0.221.
On the propane dehydrogenation example, the workflow both reproduces the established mechanism and gives a quantitative verdict on whether a metastable catalyst geometry contributes to turnover. Earlier mechanistic work on this reaction built its mechanism directly on a hand-constructed tetrahedral Pt4 without establishing that geometry's position relative to the ground state. The tetrahedral Pt4 lies 0.99 eV (95 kJ/mol) above the planar ground state, while its overall barrier along the pathways is only under 20 kJ/mol lower, so the planar geometry still governs the observed kinetics.
Perspective
The demonstrated scope is propane dehydrogenation catalyzed by a Pt4 cluster on carbon nitride, covering four reaction states and two Pt4 geometries; the authors note that converting the computed landscape into experimentally comparable kinetics requires free-energy corrections such as zero-point energy and vibrational entropy, and ultimately a microkinetic treatment that propagates the network of competing pathways into observables such as turnover frequency.
The authors explicitly do not claim state-of-the-art predictive accuracy: on Transition1x their error is 20%–60% higher than other single-model predictors, and on an isomerization dataset that contributed no training data the error is comparable to TS-GAN. The force field and DFT differ by 20–40 kJ/mol in the I2 state, so quantitative barriers still need DFT refinement, and generalization to surface systems is constrained by scarce high-quality transition-state data.