{"id":"aea70941-b5cd-4937-830a-5c52af9f6958","arxiv_id":"2507.16521","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AEFM is a learned, structure-only refinement method that iteratively improves low-fidelity transition state geometries toward DFT-quality structures.","lead":"A new deep learning method, called AEFM, refines rough transition state guesses for chemical reactions into more accurate structures without using any energy or force information. It cuts median barrier height error to about 0.79 kcal/mol and more than doubles the speed of downstream quantum optimization for one common starting point.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline accuracy metrics may conflate genuine refinement with convergence to alternative TSs: 26/1073 test cases are re-labeled, and React-OT outputs are never checked for TS validity, so the claimed median RMSD/barrier improvements may not reflect intended-reaction accuracy.","rationale":"Good-faith reading: AEFM is a practical refinement tool and the xTB validation (validity 27% to 68%, median DFT steps reduction by 10) is strong and directly tied to downstream optimization. The paper is transparent about the outlier and the re-labeling, which deserves credit. However, the headline React-OT accuracy numbers (0.088 Å, 0.79 kcal/mol) are the main quantitative claim, and they are vulnerable to the alternative-TS ambiguity because the model is not reaction-aware. The paper's own footnote shows the re-labeling changes the mean metrics; the median is robust, but the fact that 26 cases needed re-labeling signals a structural limitation. The absence of a validity/Hessian check for React-OT outputs means we cannot tell whether the median barrier error reflects true saddle points or non-stationary averages. The Gaussian prior mismatch raised by the reader is plausible but less decisive: the method is directly evaluated on actual xTB and React-OT guesses, so the prior assumption has already been tested implicitly for those two sources. The wrong-TS / re-labeling issue is more directly tied to the central claim and is not resolved by the reported numbers. A focused IRC check on a subset would settle it. If the intended-reaction match rate is high, the concern does not land and the numbers stand; if it is low, the headline claim needs qualification. The conditional verdict remains appropriate because independent reproduction and this additional validation are both needed before the numbers are taken at face value.","tokens_in":19477,"tokens_out":16747,"duration_ms":187247,"concrete_test":"Run Sella P-RFO TS optimizations at ωB97x/6-31G(d) on a random subset of 100 React-OT + AEFM outputs (same protocol as §4.4). For each converged stationary point with exactly one imaginary frequency, follow the intrinsic reaction coordinate in both directions and check whether the endpoints match the intended reactant/product pair. Report the fraction of intended-reaction matches, and recompute the Table 1 median RMSD/barrier errors after excluding the 26 re-labeled cases. If the intended-reaction match rate is materially below the valid-TS rate, or if the median errors shift by more than the reported improvements, the headline numbers overstate intended-reaction fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AEFM is a structure-only, time-independent map trained to denoise perturbed TS geometries; it receives no reactant/product context. At a fixed point it can therefore converge to a chemically valid but unintended TS. The paper documents this explicitly in the Fig. 3c outlier: after refinement the sample is 0.793 Å / 120.993 kcal/mol from the intended TS but 0.048 Å / 0.256 kcal/mol from an alternative TS. It then re-labels 26/1073 React-OT test cases in Table 1 footnote (a) whenever the alternative TS is at least 30% closer than the intended TS, changing mean RMSD from 0.188 to 0.176 Å and mean barrier error from 3.341 to 3.158 kcal/mol. The central React-OT claim (median RMSD 0.088 Å, median barrier error 0.79 kcal/mol) is therefore partly a statement about how often AEFM finds some nearby TS, not necessarily the intended one. In addition, chemical validity (exactly one imaginary frequency) is reported only for GFN2-xTB-initialized samples (27% to 68% on 100 reactions). For the React-OT samples that carry the headline numbers, no equivalent validity check is reported, so we do not know how many refined outputs are stationary saddle points rather than non-stationary averages produced by the fixed-point iteration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces Adaptive Equilibrium Flow Matching (AEFM), a structure-only refinement method that aims to improve low-fidelity transition state (TS) guesses. The model is trained to denoise Gaussian perturbations of reference TS geometries from the Transition1x dataset, with the noise scale σ determined per low-fidelity source by matching the mean RMSD of that source. At inference, the time-independent model is applied iteratively (with Anderson acceleration) to convergence, producing a refined structure. The paper reports that AEFM reduces the median RMSD of React-OT TS guesses from 0.092 Å to 0.088 Å, lowers the median absolute barrier-height error from 1.092 to 0.793 kcal/mol, increases the fraction of valid GFN2-xTB TS structures from 27% to 68% on a 100-reaction subset, and reduces the median number of DFT optimization steps by 10 for xTB-initialized searches. The authors also propose a bond-length loss that improves agreement with reference bond-length distributions.","tokens_in":19827,"tokens_out":10634,"duration_ms":110838,"significance":"The paper addresses a practical bottleneck in computational chemistry: the need to refine approximate TS guesses into DFT-quality saddle points. The proposed method is lightweight (a fraction of a second per structure), model-agnostic, and does not require energy or gradient evaluations. The authors are transparent about a key failure mode: AEFM can converge to a chemically valid but unintended TS, and they analyze this in Fig. 3c and the footnote to Table 1. The method is evaluated on a standard benchmark (Transition1x) and compared against established baselines. The main benchmark is not circular because the model is trained on Gaussian perturbations and evaluated on held-out test data. If the validity of the refined structures for the React-OT benchmark can be established, the method would be a useful contribution. However, several load-bearing claims currently lack direct support.","major_comments":[{"comment":"The chemical validity of refined outputs, defined by exactly one imaginary frequency, is reported only for GFN2-xTB-initialized structures on a 100-reaction subset. For the React-OT outputs that produce the headline numbers (median RMSD 0.088 Å, median |ΔE_TS| 0.793 kcal/mol), no analogous frequency analysis is reported. Since the barrier-height errors in Table 1 are computed as single-point energy differences at the refined geometries, and these geometries are not verified to be stationary saddle points, the reported energies may not be meaningful TS energies. Please provide frequency or stationarity checks for a representative subset of React-OT-refined structures, or clearly state that the metrics are computed on non-stationary structures and discuss the implications for the claimed high-fidelity TS geometries.","section":"Section 4.4 and Table 1"},{"comment":"The training loss is written as an expectation over x0, x1, and t, but φθ has no time input. If t is only used to define the interpolant xt via Eq. (2), this should be stated explicitly, including how t is sampled (e.g., uniform on [0,1]) and whether the expectation is over the resulting mixture distribution. More substantively, the paper claims that removing time conditioning enables the model to implicitly infer the quality of a given input, but a model that sees only the structure cannot distinguish a small isotropic perturbation from a large one. The adaptive behavior of AEFM is currently demonstrated only through the number of fixed-point iterations; no evidence is provided that the iteration count correlates with initial error magnitude. Please add such an analysis (e.g., iterations versus initial RMSD) or revise the adaptivity claims.","section":"Section 4.2.1, Eq. (10)"},{"comment":"The mean RMSD increases after AEFM for both React-OT rows (0.183 to 0.188 Å and 0.211 to 0.214 Å), and the mean barrier error for React-OT improves only from 3.405 to 3.341 kcal/mol (approximately 2%). The improvements highlighted in the abstract are median-based, and the distribution appears to be skewed by cases such as the outlier in Fig. 3c, where the barrier error increases from 17.9 to 121.0 kcal/mol. The paper should discuss the cause of the mean degradation and report the fraction of samples that are made worse by refinement, in addition to the median improvements, to support the claim of robust refinement.","section":"Section 2, Table 1"},{"comment":"The training prior models real low-fidelity errors as isotropic Gaussian noise with a source-specific σ fitted to the mean RMSD. This assumption is acknowledged in the paper, but the consequences are not fully quantified. The model must generalize from Gaussian perturbations to the structured, method-specific errors of React-OT and xTB. The paper's own outlier analysis shows that fixed-point refinement can converge to an alternative, chemically valid TS, and the footnote to Table 1 identifies 26/1073 React-OT cases where the alternative TS is at least 30% closer than the intended TS. For a method whose stated purpose is to refine guesses for a specific reaction, the frequency and characteristics of such unintended convergence should be analyzed more directly (e.g., by comparing refined structures against all nearby TSs in the dataset and reporting how often the refined sample matches an alternative TS). Please provide this analysis or explicitly frame the method as a find-any-nearby-TS tool rather than a refine-the-intended-TS tool.","section":"Section 4.2.1, Eqs. (6)-(9)"}],"minor_comments":[{"comment":"The re-labeling row should be clearly described as an exploratory post hoc analysis, distinct from the main benchmark, to avoid the impression that the main results use re-labeled targets.","section":"Table 1, footnote (a)"},{"comment":"The RMSD definition differs from that used in React-OT (normalization by 3N versus N). The paper notes this in one sentence, but it would help to provide baseline RMSD values recalculated with the same metric for a fair comparison.","section":"Section 4.5, Eq. (17)"},{"comment":"The caption and the main text disagree on which axis corresponds to the intended TS: the caption says the x-axis is the intended TS, while the text says the y-axis is the intended TS. Please reconcile this inconsistency.","section":"Figure 3c"},{"comment":"Equation (6) defines x0 = x1 + σϵ, but two lines below the text writes x0 = x1 − σϵ; the sign is inconsequential, but the inconsistency should be fixed.","section":"Section 4.2.1"},{"comment":"The phrase increases the success rate of locating valid TSs by 41% is ambiguous; the main text specifies that this is an absolute percentage-point increase from 27% to 68%, and the abstract should say so explicitly.","section":"Abstract"},{"comment":"The bond loss is defined using the ground-truth neighbor list B(x1); the paper should state explicitly that this loss is used only in training and not at inference.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is promising. However, the absence of stationarity checks on the React-OT benchmark, which carries the headline numbers, is a substantive gap that should be addressed before publication. The adaptive terminology appears overstated relative to the evidence. There is no indication of misconduct; the data appear to be reported honestly, and the authors are transparent about the alternative-TS failure mode. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper introduces AEFM, a structure-only, time-independent flow matching model that refines low-fidelity transition state guesses by iterating to a fixed point. The combination of an equilibrium flow with adaptive fixed-point inference is genuinely new compared to prior generative TS models, and the paper reports concrete gains: median barrier error for React-OT inputs drops to 0.79 kcal/mol, and for xTB inputs the fraction of chemically valid TSs (one imaginary frequency) rises from 27% to 68% with a threefold reduction in DFT optimization time.\n\nThe paper is honest about its limitations. It documents the alternative TS problem explicitly: the model can converge to a chemically valid but unintended TS, and Table 1 footnote a re-labels 26/1073 test cases when an alternative TS is 30% closer. The authors frame this as surface-walking behavior, which is fair. But it means the headline React-OT numbers are partly a statement about finding any nearby TS, not necessarily the intended one. And while validity is checked for xTB-initialized samples, no equivalent check is reported for the React-OT samples that carry the headline accuracies. That's the biggest gap.\n\nThe training prior — isotropic Gaussian noise with sigma set per source via Eq. 9 — is a simplification, but the model's ability to infer input quality from the structure itself makes it work in practice. The loss in Eq. 10 has time in the expectation but the model has no time input; that's a minor notational issue. The code and weights aren't available yet, which limits independent verification. The convergence analysis (spectral radius over iterations) is a nice touch.\n\nWho is this for? Computational chemists and ML researchers working on TS prediction or high-throughput mechanism discovery. It's not the final word, but it's a serious method with honest reporting. I'd give it a constructive referee: ask for the React-OT validity check, a clearer separation of intended vs. alternative TS metrics, and code release. If those are addressed, it's a solid contribution.\n\nVerdict: deserves a serious referee, not a desk reject.","headline":"AEFM introduces a genuinely new time-independent flow refinement scheme with real practical gains, but its headline accuracy numbers partly reflect convergence to alternative TSs, so treat the intended-reaction numbers with a grain of salt.","tokens_in":20300,"tokens_out":3055,"would_cite":true,"duration_ms":33846,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Learned equilibrium flow refines transition-state guesses to 0.088 Å.","keywords":["transition state refinement","flow matching","fixed-point inference","equivariant neural network","transition state search","reaction barrier heights","GFN2-xTB","DFT optimization"],"falsifier":"A paired test set in which low-fidelity TS guesses come from a method with systematically anisotropic errors, for example consistent stretching of the forming and breaking bonds, while AEFM is trained with the isotropic Gaussian prior; if the refiner then maps many inputs to alternative TSs or raises barrier-height error relative to the unrefined guesses, the central assumption is falsified.","tokens_in":1742,"feed_emoji":"🧪","tokens_out":4854,"duration_ms":107639,"temperature":0.7,"pith_summary":"AEFM is a structure-only refinement method that takes a low-fidelity transition-state geometry, whether from a machine-learning generative model or a tight-binding approximation, and iteratively maps it to a DFT-quality TS structure using only atomic coordinates and no energy or gradient evaluations. The paper claims this learned fixed-point flow reduces the median RMSD of React-OT predictions to 0.088 Å and the median absolute barrier-height error to 0.79 kcal/mol; for GFN2-xTB guesses it cuts the median barrier error by 59% and raises the fraction of chemically valid TS structures from 27% to 68%. If true, transition-state search becomes substantially cheaper, because refined guesses require roughly threefold fewer DFT optimization CPU hours and the refinement itself adds only fractions of a second per structure.","feed_headline":"Learned equilibrium flow sharpens TS guesses to 0.088 Å","feed_subtitle":"Structure-only refinement halves GFN2-xTB barrier errors and triples the speed of DFT transition-state optimization.","key_machinery":"The load-bearing object is an equilibrium flow field: a neural network $\\phi_\\theta(x)$ that predicts the high-fidelity TS endpoint directly from a structure, trained with variational flow matching on pairs $(x_1+\\sigma\\epsilon, x_1)$ where $x_1$ is a reference TS and $\\sigma$ is set by the source method's mean RMSD. Inference applies the fixed-point iteration $x_{k+1}=\\phi_\\theta(x_k)$ with Anderson acceleration until successive iterates differ by less than 0.01 Å RMSD, or until 100 iterations. A physics-based bond loss compares interatomic distances within a 2 Å cutoff against the reference, steering outputs toward realistic bond-length distributions, and the SE(3)-equivariant backbone makes the entire map invariant to rotation, translation, and atom-index permutation.","core_discovery":"On its own terms, the paper establishes that a time-independent flow-matching model, trained to predict the clean TS directly from a Gaussian-perturbed version of it and iterated to a fixed point at inference, can refine TS guesses from diverse sources toward the distribution of DFT-level transition states. The adaptive prior sets the perturbation scale from the mean RMSD of the source method, the SE(3)-equivariant backbone respects molecular symmetry, and a bond-length loss keeps local geometries chemically plausible. The result is a general refiner: it improves the median barrier-height error by 27% over React-OT alone and by 59% over GFN2-xTB alone, and it makes downstream DFT saddle-point optimization faster and more likely to converge to a valid TS.","pith_inferences":["Editorial inference: the isotropic Gaussian prior is the main risk; a direct test is to train on paired real low-fidelity and reference TS geometries rather than Gaussian perturbations and compare barrier errors and TS-assignment rates.","Editorial inference: because AEFM has no reactant or product context, it can legitimately converge to a different but structurally similar TS; re-ranking refined candidates against the reaction endpoints could recover the intended transition state without sacrificing the structure-only pipeline.","Editorial inference: the weak correlation between RMSD and energy change (Pearson 0.17) suggests that future refiners should explicitly optimize local bond, angle, and torsion geometry; adding such terms to the bond loss could further reduce barrier-height errors.","Editorial inference: the fixed-point formulation is a learned denoiser, so the same adaptive iteration could be applied to other low-fidelity-to-high-fidelity structure-refinement problems, though the paper only reports chemistry results."],"forward_implications":["A refiner trained with the appropriate error scale can be bolted onto various low-fidelity TS guessing methods, requiring only 2–5 model calls per structure and sub-second inference.","Refined GFN2-xTB guesses meet the 1.58 kcal/mol chemical-accuracy threshold for 57% of test reactions, up from 25% without refinement.","AEFM raises the fraction of structurally valid TSs, defined by exactly one imaginary frequency, from 27% to 68% and raises DFT TS-optimization convergence from 91% to 99%.","Fast refinement cuts the median number of DFT optimization steps by 10, a threefold reduction in CPU hours for the tested 100-reaction set.","Because refinement consumes no potential-energy-surface evaluations, it can be embedded in high-throughput reaction screening pipelines."],"supporting_citations":[{"why":"Supplies the high-fidelity DFT transition-state structures and the 9000/1073 train/test split used to evaluate all refinements.","marker":"[48]"},{"why":"The machine-learning TS generator whose predictions serve as the primary low-fidelity input and baseline that AEFM improves on.","marker":"[38]"},{"why":"The GFN2-xTB tight-binding method used to generate the second class of low-fidelity TS guesses that AEFM refines.","marker":"[49]"},{"why":"Defines the dataset split and the earlier diffusion-based TS-prediction pipeline that contextualizes the React-OT baseline.","marker":"[35]"},{"why":"Provides the flow-matching framework and optimal-transport probability paths that AEFM modifies into a time-independent equilibrium flow.","marker":"[40, 41]"},{"why":"Provides the variational flow matching perspective that justifies training the model to directly predict the clean endpoint x1.","marker":"[47]"},{"why":"The SE(3)-equivariant graph-network backbone that makes the refinement invariant to molecular rotation, translation, and permutation.","marker":"[50]"},{"why":"The Anderson-acceleration fixed-point solver used at inference to make refinement converge quickly for the majority of inputs.","marker":"[61]"},{"why":"Establishes the deep-equilibrium-model perspective and the Jacobian-spectral-radius condition used to analyze and justify the fixed-point iteration.","marker":"[59]"},{"why":"The climbing-image nudged elastic band algorithm used to generate GFN2-xTB TS guesses from reactant and product endpoints.","marker":"[14]"}],"fun_headline_variants":["AI flow refines TS guesses to 0.088 Å","Flow matching cuts TS barrier error to 0.79 kcal/mol","From GFN2-xTB, TS refiner boosts success 41%, triples DFT speed","Adaptive flow refinement accelerates DFT","Flow matching TS refiner lifts accuracy"],"cache_read_input_tokens":22400,"weakest_assumption_plain":"The load-bearing premise is that a low-fidelity TS guess behaves like a reference TS plus isotropic Gaussian noise whose scale is the source method's mean RMSD; if real errors are structured or method-specific, fixed-point refinement can settle on the wrong transition state or fail to converge, as seen in 6 of 1073 React-OT cases and 3 of 945 xTB cases in the paper's tests.","fun_headline_variants_meta":{"raw":{"variants":["AI flow refines TS guesses to 0.088 Å","Flow matching cuts TS barrier error to 0.79 kcal/mol","From GFN2-xTB, TS refiner boosts success 41%, triples DFT speed","Adaptive flow refinement accelerates DFT","Flow matching TS refiner lifts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002051,"raw_usage":{"total_tokens":7961,"prompt_tokens":895,"completion_tokens":7066,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":6983}},"tokens_in":511,"tokens_out":7066,"duration_ms":57558,"temperature":1.0,"reasoning_tokens":6983,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:07:13.393694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A paired test set in which low-fidelity TS guesses come from a method with systematically anisotropic errors, for example consistent stretching of the forming and breaking bonds, while AEFM is trained with the isotropic Gaussian prior; if the refiner then maps many inputs to alternative TSs or raises barrier-height error relative to the unrefined guesses, the central assumption is falsified.","supporting_citations":[{"cited_title":"& Winther, O","cited_arxiv_id":null,"evidence_quote":"Supplies the high-fidelity DFT transition-state structures and the 9000/1073 train/test split used to evaluate all refinements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The machine-learning TS generator whose predictions serve as the primary low-fidelity input and baseline that AEFM improves on."},{"cited_title":"& Grimme, S","cited_arxiv_id":null,"evidence_quote":"The GFN2-xTB tight-binding method used to generate the second class of low-fidelity TS guesses that AEFM refines."},{"cited_title":"& Kulik, H","cited_arxiv_id":null,"evidence_quote":"Defines the dataset split and the earlier diffusion-based TS-prediction pipeline that contextualizes the React-OT baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the variational flow matching perspective that justifies training the model to directly predict the clean endpoint x1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The SE(3)-equivariant graph-network backbone that makes the refinement invariant to molecular rotation, translation, and permutation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Anderson-acceleration fixed-point solver used at inference to make refinement converge quickly for the majority of inputs."},{"cited_title":"& Kolter, Z","cited_arxiv_id":null,"evidence_quote":"Establishes the deep-equilibrium-model perspective and the Jacobian-spectral-radius condition used to analyze and justify the fixed-point iteration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The climbing-image nudged elastic band algorithm used to generate GFN2-xTB TS guesses from reactant and product endpoints."}],"review_version":1}