Replaying Solved Transition States Across Mechanistically Isomorphic Reactions

Finding the transition state of a chemical reaction is the most expensive routine step in computational chemistry. The transition state is the highest-energy point a reaction must pass through, and its height sets the reaction rate. Locating it with density functional theory (DFT, the standard quantum-mechanical method for computing molecular energies) usually means running a nudged elastic band calculation. That method bends a chain of intermediate molecular snapshots until it traces the lowest-energy path from reactants to products, and it costs hundreds of DFT force evaluations per reaction. Much of this effort is repeated. A large reaction network contains many steps that are "the same move" applied to different molecules: the same bonds break, the same electrons shift, and only the spectator substituents differ. This note proposes treating each solved transition-state search the way theorem provers treat a solved proof. We lift it to a template by replacing the concrete spectator parts with variables. When a new reaction matches that template, we map the old solution onto it and use it as the starting path, so the new search begins near the answer.

The core claim is that a lifted mechanism key predicts when a solved reaction path will transfer to a new reaction, and that it does so better than similarity of molecules or a learned potential alone. The key combines the reaction center with the pattern of electron flow, and it abstracts away the substituents in the manner of second-order logic. If the claim holds, the number of DFT force calls per converged transition state should fall substantially, and the fall should be largest where learned potentials are weakest, on reactions unlike anything in their training data.

Where This Sits

Proof reuse through second-order matching supplies the method. Kolbe et al. abstract the skeleton of an inductive proof by replacing concrete function symbols with variables, then replay it in a new domain through second-order matching. Davis et al. and Kolobov et al. lift relational rules to predicate-variable templates in the same spirit, but they use the templates as a statistical bias on search, not as a solution to replay. Harandi et al. show that replaying a derivation history can beat transforming the finished artifact. Our analogue replays the optimized reaction path rather than merely reusing a reaction label. All of these systems work over discrete symbols. Ours has to carry a continuous geometry across the match, and that is the new technical burden.

Chemistry tools that reason over molecular graphs motivate the work and also constrain it. RMG v3.0 uses "loose isomorphism," which compares atom connectivity while ignoring radicals, lone pairs and charges, to speed up duplicate detection. It relies on hand-written reaction families and estimates rates rather than computing transition states. Dewyer et al. prune a generated network by graph distance and then run DFT transition-state searches only on the survivors, 36 of them for cobalt-catalyzed hydroformylation. Each of those 36 searches still starts from scratch. Both papers note the same weakness: connectivity alone cannot tell apart species that share bonds but differ electronically. Zheng et al. make the parallel point for circuits, where plain structural matching fell to roughly chance once logic was restructured.

These results shape our design in two ways. The match key must include electron flow, so it is not loose isomorphism. Matching must also go past raw topology. We therefore take the electron-flow annotation from PMechRP's ArrowFinder, which reconstructs ground-truth arrow-pushing in 99.55% of test cases, and from the bond-electron matrix representation of FlowER. Neither of those systems computes energies. We use them only to describe the mechanism, and DFT does the physics. The approach of Ulissi et al. does connect to DFT: it cuts transition-state calculations by 95% with scaling relations and active learning. However, it estimates barriers instead of computing paths, and it treats each surface network in isolation.

Research Program

The first component is a mechanism key. For each reaction we build its atom-mapped bond-electron change: which bonds form and break, and which lone pairs and radicals move. We keep the reacting atoms and a shell of neighbors around them concrete. Everything else is replaced by typed variables that stand for any substituent group attached at that position. The size of the concrete shell is the abstraction level.

We pick that level by compression over the library of solved reactions, in the manner of STEVIE. The chosen templates minimize the total description length of the library under a penalty on vacuous templates. STEVIE's exact solver becomes intractable above about 460 literals. To stay clear of that limit, we run compression only within buckets that share an identical reaction center, which keeps each problem small.

The second component is geometric replay. When a new reaction matches a template, the second-order match gives an atom correspondence. We align the stored path for the matched atoms by a rigid Kabsch fit frame by frame. Unmatched substituent atoms are placed by constrained relaxation with a cheap force field while the matched core is held fixed. The result becomes the initial band for a climbing-image nudged elastic band run at the same level of theory as the source.

The third component is a verification loop, which keeps bad matches from being silently accepted. A replayed band that fails to converge within a budget falls back to a standard initialization. The final saddle point is checked for exactly one imaginary frequency along the intended bond changes. Every false match is logged, so we can measure how often the template key over-generalizes. This is the chemical counterpart of negative transfer, a problem the logic-transfer literature leaves open.

Baselines

The do-nothing reference is standard band initialization with no prior knowledge: linear or image-dependent pair-potential interpolation between the relaxed reactant and product, run with the same optimizer, convergence threshold and force-call cap as every other arm. We found no reported result on Lacuna for force-call counts of this initialization on Transition1x, so we will measure it ourselves.

The strongest existing competitor is pre-optimizing the band with a reactive machine-learning potential and then refining it with DFT. We will use PaiNN, an equivariant neural network potential, trained on the training split of Transition1x. That paper reports a barrier mean absolute error of 0.075 eV for PaiNN trained on Transition1x, against 0.495 eV when trained on the equilibrium-only ANI1x data. It is the strongest competitor because a potential this accurate on barriers may already place the band close enough that replay adds nothing. It gets exactly the same training reactions that populate our replay library, the same DFT budget, and a matched hyperparameter search.

The retrieval control is the ablation that isolates the key ingredient. It keeps our geometric replay but chooses the source reaction by reactant-plus-product fingerprint similarity instead of by the lifted mechanism key. A second ablation uses a ground-level key: the exact reaction center at a fixed radius, with no variable lifting and no electron-flow term. This tests whether lifting and electron flow matter separately. The final arm combines replay and the potential, relaxing the replayed band with PaiNN before DFT, to test whether the two are complementary.

Replay beats the competitors if it reduces median DFT force calls to convergence relative to PaiNN pre-optimization while matching or exceeding convergence rate and barrier fidelity. We define barrier fidelity as agreement within 0.05 eV of the reference barrier with the same imaginary mode. The result should also hold on held-out splits.

Why not just…?

Why not just use the reactive machine-learning potential? This is the real threat. On in-distribution reactions, a potential with 0.075 eV barrier error may leave replay no room to help. Our bet is out of distribution. The Transition1x paper itself shows how sharply these potentials degrade on configurations they never saw: 0.495 eV versus 0.075 eV. Replay needs only one solved analogue, not dense coverage. The splits below are designed to settle this question. If PaiNN matches replay on the template-held-out and size-held-out splits too, the standalone claim fails, and we report replay only as a cheap add-on to the combined arm.

Why not just retrieve the most similar solved reaction by fingerprint? Fingerprint similarity rewards shared spectator groups. What predicts transfer, we claim, is a shared reaction center and electron movement. This is the structural gap Zheng et al. found in circuits, where surface structure and function diverge. The retrieval ablation settles the question directly. If it does as well as the lifted key, the second-order framing adds nothing, and we will say so.

Why not skip the path and just predict the barrier with scaling relations or a network? For screening, that is often enough, and Ulissi et al. show how far it goes. Even so, that approach could not rule out methanol as a product at 90% confidence because of uncertainty in the estimates. Settling such cases requires an actual converged saddle point. Replay targets exactly the calculations that active-learning pipelines still have to run.

Proposed Experiments

The main testbed is Transition1x: about 10,000 hydrogen, carbon, nitrogen and oxygen reactions with reference climbing-image paths at ωB97x/6-31G(d). We rerun that functional so costs are directly comparable. We use three splits. The random split measures in-distribution performance. The template-held-out split puts reaction centers in the test set that never occur in training at the concrete level but do occur after lifting, which is where lifting should earn its keep. The size-held-out split trains on reactions with up to five heavy atoms and tests on reactions with six or seven, which is where learned potentials should struggle.

The metrics are median and 90th-percentile DFT force calls to convergence, convergence rate under a fixed cap, barrier fidelity, and false-match rate. The decisive check is the template-held-out split. The lifted key must beat both fingerprint retrieval and PaiNN pre-optimization on force calls without losing fidelity. If it does not, the central hypothesis is wrong.

A secondary case study reuses within a single network. We take the cobalt hydroformylation network of Dewyer et al. and order its 36 transition-state searches so that later steps sharing a lifted template with earlier ones are seeded by replay. We then report total DFT cost against their from-scratch protocol. Because that network involves spin and coordination subtleties at the metal, it also tests whether the electron-flow key avoids false matches that connectivity alone would make. False matches here would show up as saddle points with the wrong imaginary mode.