acceptodds
Under review as a conference paper at ICLR 2027

When Noise Helps: Probabilistic Kinetic Solving for Reaction Networks

Abstract

Predicting the outcome of a chemical reaction from first principles would let us screen molecules, materials and drugs on the computer instead of in the lab. Reaction networks, which map out the possible pathways a reaction can take, are a standard tool for this. A kinetic solver turns a network into a reaction outcome, but its reaction rates depend exponentially on the energy barriers computed by quantum chemistry. Even small barrier errors, within reach of the most expensive methods, can flip the predicted product. We introduce a kinetic solver that treats barriers as distributions instead of fixed numbers, and predicts the expected outcome over that uncertainty. This gives more accurate and more reproducible predictions, at a cost that shrinks as uncertainty grows. We validate these gains on synthetic networks spanning a wide range of topologies, temperatures, and reaction times, and qualitatively on real reaction networks from the literature. As machine learning pushes reaction network exploration towards full automation, manual error curation becomes infeasible, and uncertainty aware solvers offer a path to reliable chemistry prediction at scale.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty quantification, reaction networks, AI for science, Monte Carlo Estimation, probabilistic prediction, chemical kinetics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HxEiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks what to do when the barriers in a reaction network are uncertain. Instead of propagating one set of estimated barriers, it averages the predicted populations over perturbed barriers. The perturbations are generated from shared minimum and transition-state energies, so detailed balance is preserved for each sample. The paper then uses the assumed uncertainty to decide when a Krylov solve has done enough work. Results are reported on synthetic networks and on two literature-derived ozonolysis networks.

Critical strengths
  1. The main idea is interesting. The paper uses the same uncertainty model twice: first to average the predicted populations, and then to decide how much numerical accuracy is worth paying for. That is a sensible connection between uncertainty in the input and the cost of solving the dynamics. The shared-energy construction also avoids breaking detailed balance by adding unrelated noise to each directed edge.

  2. The controlled experiments are consistent with the proposed mechanism. On the synthetic networks, the uncertainty-aware predictor generally has lower population RMSE and better top-K accuracy than the deterministic baseline. The time-temperature sweep and the runtime study make the evaluation more useful than a single toy example. The CRN-E failure is also informative: the method can reweight paths that are present, but it cannot recover a reaction that is missing from the network.

Critical weaknesses
  1. The evidence is still mostly tied to the assumed synthetic error model. The quantitative experiments generate the barrier errors with the same Gaussian covariance structure used by the method, and most of them set the assumed uncertainty equal to the true uncertainty. Figure 2b changes the assumed scale, which is useful, but it does not test biased errors or a different correlation structure. The two real-network examples are qualitative. CRN-T moves the prediction in the right direction but still misses the reported yield, while CRN-E cannot be solved because a required pathway is absent. The real calculations also omit free-energy corrections, and the choice of 4 kcal/mol is not calibrated against an independent reference. I would therefore describe the results as a promising first demonstration, not yet as evidence that the method is reliable for realistic reaction networks.

  2. The truncation rule needs a clearer error story. Appendix B derives the eigenvalue-variance relation for independent edge perturbations, but the method in Section 4.1 uses correlated perturbations induced by shared minima and transition states. That derivation does not directly cover the model used in the experiments. Also, a small incremental SNR in Eq. (10) does not by itself tell us how large the remaining error in the ensemble mean can be. Figure 3 is a useful sanity check, but the paper would be stronger if it reported the discrepancy between the truncated solver and a converged ensemble calculation across the main experiments. A comparison with an ordinary tolerance-based stopping rule at matched accuracy would also help.

Questions for authors
  1. How much do the gains change if the barrier errors have a systematic bias or a different covariance structure? For CRN-T, how was the 4 kcal/mol uncertainty chosen, and does the qualitative improvement persist over a reasonable range of values?

  2. Can you report the error between the truncated and converged ensemble predictions using the same barrier samples? It would be useful to see representative worst-case or upper-quantile errors, as well as a comparison with tolerance-based stopping at matched prediction accuracy and hardware.

  3. Is Appendix B intended to apply to the correlated perturbations used in Section 4.1? More generally, should Eq. (10) be read as an empirical stopping heuristic, or is it meant to provide a numerical error guarantee? These points are important for interpreting the scope of the claims. I do not think the paper needs a large new benchmark for the current submission, but the authors should be careful not to present the synthetic results as a validation of realistic chemistry errors.

1 comment
AuthorsOpenReview ↗

Response to Reviewer HxEi

We thank the reviewer for the constructive feedback and address each point below.

The quantitative experiments generate the barrier errors with the same Gaussian covariance structure used by the method, and most of them set the assumed uncertainty equal to the true uncertainty. Figure 2b changes the assumed scale, which is useful, but it does not test biased errors or a different correlation structure.

As the reviewer notes, Fig. 2b varies only the assumed scale σ'. We have extended this into a sweep over σ and σ′ (Appendix E.1): p̄ improves over p for all 0 < σ′ ≤ 2σ, for σ ∈ {3, 6, 12} kcal/mol. p̄ therefore needs only a rough estimate of the barrier uncertainty, not a calibrated one. To test errors outside p̄'s noise model, we added an experiment in which the true barrier errors carry a systematic bias that p̄'s zero-mean model does not include. q̄ still reduces the total error across the whole bias range [−12, 12] kcal/mol for multiple σ ∈ {3, 6, 12} kcal/mol (Appendix E.2)

The two real-network examples are qualitative. CRN-T moves the prediction in the right direction but still misses the reported yield. The choice of 4 kcal/mol is not calibrated against an independent reference.

We now report the yield for CRN-T. The experimental rate constant (Dowideit & von Sonntag, 1998) implies at least 57% conversion after 1 ms; q̄ predicts 23%, against 0.5% for p. We also swept σ' ∈ {2, 4, 6, 8} kcal/mol, and the improvement of q̄ over p persists (Appendix E.3). As this is a single example, we still regard it as qualitative. The quantitative evaluation is on synthetic networks (Section 5.1–5.2).

Appendix B derives the eigenvalue-variance relation for independent edge perturbations, but the method in Section 4.1 uses correlated perturbations... Is Appendix B intended to apply to the correlated perturbations used in Section 4.1? ... should Eq. (10) be read as an empirical stopping heuristic, or is it meant to provide a numerical error guarantee?

Appendix B is not intended to cover the correlated perturbations of Section 4.1, and the SNR rule (Eq. 10) does not depend on it: Eq. 10 is an empirical stopping rule computed directly from the sampled populations, not a numerical error guarantee. Appendix B supports only the intuition in Section 4.2 that fast modes are the most affected by noise. We now state its role explicitly.

Figure 3 is a useful sanity check, but the paper would be stronger if it reported the discrepancy between the truncated solver and a converged ensemble calculation across the main experiments.

We thank the reviewer for this point. q̄ and p̄ reach the same population RMSE in all experiments in Sections 5.1-5.3. This was stated only briefly at the start of Section 5. We have now quantified this in Section 5.1: with the same networks and barrier samples, their population RMSEs differ by at most 1e-4 at every M and σ. The SNR cutoff therefore adds no measurable error, and the speed-up comes at matched accuracy. We now state this clearly in experiment 5.1

Reviewer_ZSKLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses forward kinetics on chemical reaction networks whose activation free energies ΔG‡ come from approximate electronic-structure methods and therefore carry errors of 1–5 kcal/mol, which Eyring kinetics amplifies exponentially. It proposes:

  1. Replacing the deterministic population p(ΔĜ‡) = exp(tK(ΔĜ‡)) p_init with the expected population p̄(ΔĜ‡, σ′) = E_{ε′ ~ N(0, Σ_σ′)}[ p(ΔĜ‡ + ε′) ], where ε′ is a correlated, detailed-balance-preserving Gaussian perturbation (per-node η_i, per-transition-state η‡_ij, variance split α), estimated with L = 128 Sobol′ Monte Carlo samples (§4.1, Eqs. 5–7).
  2. A cheaper surrogate q̄ in which each sample's matrix exponential is approximated by shift-and-invert Arnoldi, with the Krylov rank chosen by an SNR rule that stops when the mean increment across samples falls below its across-sample spread (§4.2, Eq. 10), motivated by a first-order argument (App. B) that eigenvalue uncertainty scales with eigenvalue magnitude.
  3. Validation on synthetic chemistry-mimicking networks (M ∈ {10, 100, 1000}; sweeps over σ, T, t) showing lower population RMSE and marginally higher top-K accuracy than the plug-in solver; 1.3–2.3× speed-up of q̄ over naive MC; and two Chemoton ozonolysis networks, one where q̄ recovers the experimental product (CRN-T) and one where both methods fail because a step is missing (CRN-E).
Critical strengths
  1. The uncertainty-to-rank coupling is a promising direction . applying the Krylov truncation to the epistemic floor in stop resolving spectral modes that barrier uncertainty will wash out anyway which the central idea of the paper. The principle that numerical effort should scale down with parameter uncertainty is sound and under-exploited in kinetics solvers.

  2. Clean, systematic empirical design . Sweeps over σ, M, T and t with 10³–10⁴ networks per point; a bias–variance decomposition (Fig. 2, App. A); a σ′-misspecification curve (Fig. 2b) showing over-estimation is safe; timings; and an honest negative real-world case (CRN-E). Presentation is clear and Fig. 1 motivates the problem well.

Critical weaknesses

W1. The central empirical claim is a known theorem, stated nowhere, and tested only where it is guaranteed to hold

  • Under a flat prior on ΔG‡, p̄(ΔĜ‡, σ) = E[p(ΔG‡) | ΔĜ‡] by Gaussian symmetry — the posterior mean, which minimizes Bayes risk under squared loss. The RMSE advantage in Figs. 2a and 4 therefore follows a priori whenever the truth is corrupted with the predictor's own noise model.
  • That is exactly the experimental design: same Gaussian family, same Σ construction, same α = ½, σ′ = σ. The only misspecification ever tested is the scale σ′ (Fig. 2b).
  • The robustness claim ("lower variance under perturbations", l.053) is the O(1/σ′) Lipschitz bound for Gaussian smoothing of a [0,1]-valued function (randomized smoothing: Cohen, Rosenfeld & Kolter, ICML 2019; Salman et al., NeurIPS 2019; Nesterov & Spokoiny, FoCM 2017). The bias–variance "mechanism" (ll.241–242) is Bishop (1995) / SIMEX (Cook & Stefanski, JASA 1994). None is stated or cited.
  • Untested: noise shape, α, and above all bias — yet the conclusion concedes real errors are "systematic and correlated" (ll.483–484). A symmetric smoother cannot remove bias.
  • Bayes-optimality is an average statement over the prior; a network sitting at a kinetic threshold can lose. No per-network win-rate or gain distribution is reported.

W2. The one real-world positive result is explained by a generic artefact of the method

Under Gaussian barrier noise, rates are log-normal: E[k̃] = k · exp(s²/2), s = σ′/k_BT. At the CRN-T setting (σ′ = 4 kcal/mol, 298 K), s ≈ 6.75 and the factor is ≈ 10¹⁰, so symmetric smoothing pulls any sub-threshold plug-in prediction up and earlier. CRN-T — "p rises two orders of magnitude too late; q̄ reaches conversion in the experimental window" — is precisely this direction, with σ′ hand-set post hoc (ll.432–433) and no sensitivity sweep. The error sources named to justify σ′ = 4 (missing thermal corrections, implicit solvation) are systematic, not zero-mean Gaussian. A discriminating validation needs the opposite sign (a case where p is too fast and q̄ corrects it) or a σ′-sweep showing the 23% is not an artefact of the chosen σ′.

W3. The new contribution (SNR truncation) rests on an analysis of a different model and a heuristic without a bound

  • App. B analyses a model the algorithm never runs. Eq. 12 assumes independent per-directed-edge noise, which breaks detailed balance and contradicts the correlated construction of §4.1. The §4.2 mode-decay argument relies on real eigenvalues, which independent noise does not guarantee.
  • Eq. 14 is a non sequitur. Only the lower bound r_i ≥ 1/√(M(M−1)) is proved (Cauchy–Schwarz). r_i = ‖g^(i)‖₂ / |Σ_ab g^(i)_ab| is unbounded above whenever |λ_i| is small by cancellation, so slow modes can be the most uncertain in relative terms; "absolute uncertainty ∝ |λ_i|" does not follow. What is provable: √Var(δλ_i) ≥ (σ′/k_BT) |λ_i| / √(M(M−1)).
  • Eq. 10 is a design choice, not a convergence test. It compares the mean increment to the per-sample spread rather than the standard error (√L ≈ 11× smaller): "SNR < 1" means "below the epistemic floor", not "statistically zero". No bound on ‖q̄ − p̄‖ follows, and stopping at the first crossing presumes monotone Krylov increments.
  • Direct analogues exist. The statistic ‖mean increment‖²/variance is the norm test of Byrd, Chin, Nocedal & Wu (Math. Prog. 2012) / inner-product test (Bollapragada, Byrd & Nocedal 2018); "resolve numerical error only to the level of other uncertainty" is the probabilistic-numerics principle (Hennig, Osborne & Girolami, Proc. R. Soc. A 2015; Wenger & Hennig, NeurIPS 2020); coarse solves for most samples is multilevel MC (Giles, Oper. Res. 2008) — and Döpking's multilevel sparse grids in this exact application.
  • Measured payoff is modest: 1.3–2.3× over naive MC (l.402); ~7× slower than the deterministic solver it replaces (117 ms vs 16 ms, M = 1000). The "7× rather than 128×" is 16-core parallelism, not the rule.

W4. The main contribution is a Gaussian-smoothed (posterior-mean) estimator that is standard in both the ML/statistics and the chemical-kinetics UQ literatures, and the paper neither identifies it as such nor compares against anything.

  • Eq. 6 is the Gaussian convolution p ∗ φ_Σ is the Weierstrass transform and at σ′ = σ it is exactly the posterior mean E[p(ΔG‡) | ΔĜ‡] under a flat prior (one line, Appendix A.1 of this review). As an estimator this is randomized smoothing (Cohen, Rosenfeld & Kolter, ICML 2019; Salman et al., NeurIPS 2019; Yang et al., ICML 2020 on kernel choice), Gaussian smoothing for non-smooth objectives (Nesterov & Spokoiny, FoCM 2017), training-with-noise ≡ Tikhonov (Bishop, 1995), the simulation step of SIMEX (Cook & Stefanski, JASA 1994), and GP prediction under input noise (Girard et al., NIPS 2003). None is cited.

Minor weaknesses

  • Title / abstract : nothing about noise helps — integration over the barrier posterior helps, and a deterministic quadrature (sparse grids, sigma points) yields the same estimator. "Cost shrinks as uncertainty grows" is shown only as stop-rank (Fig. 3c), never as wall-clock vs σ′; all timings are at σ′ = 6.
  • "q̄ ≈ p̄ everywhere we tested" rests on Fig. 3 alone (M = 100, T = 1000 K, t = 0.01 s); surrogate accuracy is never reported on the §5.1–5.2 grids where the headline gains are measured.
  • T-dependence : the gain increasing with T contradicts the paper's own Fig. 1a logic (exp(σ/k_BT) shrinks with T). The absolute gain RMSE(p) − RMSE(q̄) likely conflates "how much the method helps" with "how much chemistry happens" at low T / short t.
  • Top-K : q̄ is the wrong Bayes rule for a 0–1 loss; argmax_i P(i dominant | ΔĜ‡) — the fraction of MC samples in which i has the largest population — is free from the samples already drawn and is what Medford et al. 2014 compute. Top-K gains are marginal and overlap at M = 10.
  • "1/L rate in RMSE" : the estimator converges at L^(−1/2); what decays as 1/L is the excess over the irreducible floor. The QMC advantage in d ≈ M + |E|/2 ≈ 170 dimensions with L ≤ 512 is asserted, not shown.
  • Synthetic prior : topology is matched to real networks (App. D) but the barrier/energy distribution — which governs the gain, since at 400 K / 1 h the reactive threshold is ≈ 30 kcal/mol (k·t = 1 ⇔ ΔG‡ ≈ 37.9 k_BT) and barriers are drawn uniformly up to 46 kcal/mol above the higher endpoint — is not.
  • Cost accounting: the per-sample sparse LU of (I − tK̃_ℓ) (Alg. 2, line 1) is untouched by rank truncation and plausibly dominates; dense expm is a straw-man baseline when the paper's own Krylov p exists.
  • Smaller: notation flip between Eq. 2 (K_ij = k_ji) and App. B (K_ab = rate b→a); stray "." in the Fig. 1 caption; the clip-and-renormalize in Alg. 2 line 19 (which makes q̄ biased and is what keeps Eq. 7's ½ meaningful) belongs in the main text; no code is mentioned; under detailed balance K is symmetrizable, so Lanczos would do.
Questions for authors

Clarification

  1. Do you agree that p̄(ΔĜ‡, σ) is the posterior mean under a flat prior, so that the Fig. 2a / Fig. 4 RMSE advantage is guaranteed under the well-specified design? If so, what does the synthetic suite test beyond that?
  2. Results under biased noise ε ~ N(μ, σ²) with μ < 0 (barriers systematically under-estimated), and under misspecified α — does q̄ still beat p?
  3. The per-network distribution and win-rate of RMSE(p) − RMSE(q̄), not only the mean over 10⁴ networks.
  4. CRN-T: sensitivity to σ′ ∈ {2, 4, 8}; and is there any real network where p is too fast and q̄ corrects it?
  5. ‖q̄ − p̄‖ (or RMSE(q̄) beside RMSE(p̄)) on the §5.1 and §5.2 grids; wall-clock vs σ′ at fixed M; LU-vs-Arnoldi cost breakdown.
  6. App. B under the correlated noise model of §4.1 — does the relative-uncertainty claim survive, given that r_i is bounded only from below?
  7. Relative gain (or RMSE(p) alongside) in Fig. 5, to resolve the T-dependence.

Suggestions

  • State the Bayes / Gaussian-smoothing argument explicitly, with two or three references (randomized smoothing, Nesterov–Spokoiny, SIMEX).
  • Report quantiles of the sample distribution, not only its mean.
  • Use posterior dominance (fraction of samples) for top-K.
  • Compare the SNR rule against multilevel MC and a sparse-grid quadrature.
  • Lanczos instead of Arnoldi under detailed balance.

===============

Sources

  • Sutton, Guo, Katsoulakis, Vlachos — Effects of correlated parameters and uncertainty in electronic-structure-based chemical kinetic modelling, Nat. Chem. 2016
  • Medford et al. — Assessing the reliability of calculated catalytic ammonia synthesis rates, Science 345:197 (2014)
  • Döpking, Plaisance, Strobusch, Reuter, Scheurer, Matera — Addressing global uncertainty and sensitivity in first-principles based microkinetic models by an adaptive sparse grid approach, J. Chem. Phys. 148:034102 (2018)
  • Reagan, Najm, Ghanem, Knio — Uncertainty quantification in reacting-flow simulations through non-intrusive spectral projection, Combust. Flame 132:545 (2003)
  • Proppe, Husch, Simm, Reiher — Uncertainty quantification for quantum chemical models of complex reaction networks, Faraday Discuss. 195:497 (2016)
  • Proppe & Reiher — Mechanism deduction from noisy chemical reaction networks (KiNetX / AutoNetGen), JCTC 15:357 (2019); arXiv:1803.09346

More references (Cohen et al. 2019; Salman et al. 2019; Nesterov & Spokoiny 2017; Bishop 1995; Cook & Stefanski 1994; Byrd et al. 2012; Bollapragada et al. 2018; Hennig et al. 2015; Wenger & Hennig 2020; Gavish & Donoho 2014; Giles 2008; Berthet et al. 2020) are standard but were not re-fetched here — spot-check exact titles/venues before citing.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ZSKL

We thank the reviewer for the detailed feedback. We address each point raised by the reviewer below. A summary of the new experiments is given in our general comment to all reviewers.

W1. (a) Known theorem, not stated.

We thank the reviewer for pointing out this connection, which we now state in Section 4.1: under a flat prior and Gaussian errors with σ′ = σ, p̄ is the posterior mean of the populations and therefore the MMSE estimator under that prior. We see this as theoretical support for the method rather than a weakness. In our setting, however, it is not an exact guarantee, since the barriers are drawn from bounded distributions. It also does not predict how large the gain is across network size, time and temperature (Figures. 4-5), or the cost reduction from SNR truncation (Section 4.2).

W1(b) Only σ′ misspecification tested.

We have extended this into a sweep over the σ × σ′ grid (Appendix E.1). p̄ improves over p for all 0 < σ′ ≤ 2σ, for σ ∈ {3, 6, 12} kcal/mol, showing a rough, conservative estimate of σ is enough. To test errors outside p̄'s noise model, we also added an experiment with systematic barrier bias, described in W1(d).

W1(c) Robustness / bias–variance mechanism not cited. (randomized smoothing: Cohen, Rosenfeld & Kolter, ICML 2019; Salman et al., NeurIPS 2019; Nesterov & Spokoiny, FoCM 2017). The bias–variance "mechanism" (ll.241–242) is Bishop (1995) / SIMEX (Cook & Stefanski, JASA 1994)

We thank the reviewer for these references and have added the most relevant ones to the revision in the Related Work (Section 6). Following the review's note to "spot-check exact titles/venues before citing", we verified all references before adding them.

W1 d) Untested: noise shape, α, and above all bias — yet the conclusion concedes real errors are "systematic and correlated" (ll.483–484). A symmetric smoother cannot remove bias.

We agree, and we do not claim that q̄ removes bias. Our claim is that it reduces the variance part of the error, which is why bias is listed as a limitation in Conclusions Section. Bias is usually specific to an electronic structure method and can commonly be reduced by calibrating against reference data. We agree, that it is important to check that the variance reduction holds when bias is present. We therefore added an experiment (Appendix E.2) in which the barrier errors include a systematic bias b that p̄ does not account for. For b ∈ [−12, 12] kcal/mol and σ ∈ {3, 6, 12} kcal/mol, p̄ still achieves 8-24% lower population RMSE than p.

W2. The one real-world positive result is explained by a generic artefact of the method

q̄ is the expected population under the stated barrier uncertainty, so the shift described here as an artefact is what the method is designed to do. As the reviewer suggests, we added a σ′ sweep from 2 to 8 kcal/mol on CRN-T (Appendix E.3). It shows that the gain over p is consistent across σ′ and is not specific to σ′ = 4 kcal/mol. We agree that CRN-T is a single qualitative example. To our knowledge, no large public set of real networks with measured outcomes exists, which is why our quantitative evaluation uses synthetic networks.

W3. The new contribution (SNR truncation) rests on an analysis of a different model and a heuristic without a bound.

The SNR truncation rule does not depend on the Appendix B analysis. The rule (Eq. 10) is computed directly from the sampled populations, and its accuracy and speed-up are measured empirically in Section 5. Appendix B only motivates one intuitive sentence in Section 4.2. We thank the reviewer for bringing this to our attention, as it suggests our text was unclear. We now:

  • State explicitly that Appendix B provides intuition for the SNR criterion, not a guarantee.
  • Show that r_i is bounded in practice in Appendix B.

W4. The main contribution is a Gaussian-smoothed (posterior-mean) estimator that is standard in both the ML/statistics and the chemical-kinetics UQ literatures, and the paper neither identifies it as such nor compares against anything.

We now cite some of the most relevant works listed by the reviewer (Section 4.1). They apply Gaussian smoothing to different problems (classifier robustness, optimisation, regression) and do not address kinetic solvers, so they are not direct baselines in our setting. In chemical kinetics, uncertainty propagation is typically used to report error bars on predictions (Proppe et al., 2016; Proppe & Reiher, 2019). We instead show that averaging over barrier uncertainty improves the prediction itself. Against the standard deterministic solver p, our baseline throughout Section 5, q̄ gives, in ~100 ms of solver time, the accuracy gain that would otherwise take days of extra electronic-structure computation (Fig. 4, Table 1)

Reviewer_A2a2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses a practical problem in predicting reaction outcomes from first principles. Kinetic solvers turn a reaction network into predicted populations, but the reaction rates depend exponentially on activation barriers computed with approximate quantum-chemistry methods, so errors of a few kcal/mol can change the predicted outcome. Instead of treating the computed barriers as exact, the authors treat them as uncertain: they sample many physically consistent sets of barriers, run the solver on each, and report the average population. They also make this cheaper with a Krylov-based solver that stops early once the contribution of a new mode is smaller than its spread across samples. The method is evaluated on synthetic networks of 10–1000 nodes over a range of temperatures, times and noise levels, with runtime and sampling-convergence studies, and qualitatively on two published ozonolysis networks. In one, the averaged solver predicts the experimentally observed product where the standard solver predicts none; in the other, both fail because a reaction step is missing.

Critical strengths

An important and well-motivated problem.

The paper convincingly illustrates how small errors in computed barriers can substantially change kinetic predictions (§3, Fig. 1). This becomes increasingly relevant as reaction networks are generated automatically and manual inspection of barriers is no longer feasible. The problem and motivation are clearly explained.

A broad and consistent synthetic evaluation.

The authors evaluate many network sizes, temperatures, reaction times and noise levels, together with sampling-convergence and runtime studies. Within the tested setting, the averaged predictor did not perform worse than the deterministic solver (Fig. 5) and substantially reduced the sensitivity of predictions to noisy barriers (Fig. 2).

Critical weaknesses
  • The improvements are demonstrated mainly under the same error model used to build the method. In the synthetic experiments, barrier errors are generated from the same Gaussian model used for the averaging, generally with the assumed noise level equal to the true one. This is a favourable setting, and it does not show how the method behaves when the error model is wrong. Errors from approximate quantum-chemistry methods can include systematic biases and correlations across related reactions. Averaging over perturbations around a biased estimate does not correct the bias, and whether it still improves the predicted populations in that case is not tested. The paper acknowledges this limitation (§7) but does not investigate it experimentally. The real-network evidence is also limited. On CRN-T, the averaged solver assigns some population to the observed product, but this does not show that the predicted populations are quantitatively accurate, and the assumed uncertainty (σ′ = 4 kcal/mol) is chosen rather than calibrated. The gains are also modest at lower noise levels (Fig. 5), and top-1 accuracy changes little (Fig. 4). Since the paper’s motivation is more reliable predictions from approximate quantum chemistry, the lack of evaluation under realistic or misspecified errors is my main concern.

  • The uncertainty-dependent stopping rule is a distinct numerical contribution, but its practical value is unclear. It gives a 1.3–2.3× speed-up relative to full Monte Carlo propagation (§5.3), but it is not compared with standard tolerance-based Krylov stopping at matched accuracy, so it is unclear how much of the saving comes from accounting for physical uncertainty rather than from accepting a coarser approximation. At 1,000 states, the averaged solver takes about 117 ms compared with 16 ms for the deterministic solver, while computing the barriers themselves takes far longer (Table 1). The paper does not show a workflow in which this saving matters in practice, for example a setting where the solver is run many times. The uncertainty information might be more valuable for deciding which barriers need more accurate calculations, but this is not explored.

Questions for authors

Clarifications

  • Misspecified errors. How do the gains change when the true barrier errors differ from the assumed model, e.g. a systematic bias, reaction-type-dependent errors, a different correlation structure, or heavy-tailed errors? An evaluation using barriers from a cheaper method scored against a more accurate reference would be particularly informative.
  • Real-network validation. On CRN-T, how does the predicted population of the observed product change as σ′ varies between 1 and 8 kcal/mol, and how does it compare with the experimental yield? Would a simple systematic shift of the barriers produce the same qualitative result with the deterministic solver?
  • Reference baselines. How do both predictors compare with simple baselines, e.g. retaining the initial reactant population or using equilibrium populations, on the synthetic benchmarks? This would help interpret the reported RMSE values.

Suggestions

  • Since computing barriers dominates the workflow cost, it would be interesting to test whether the uncertainty information helps decide which barriers to recompute, e.g. by replacing selected noisy barriers with their true values compared with random selection.
  • The paper is motivated by first-principles prediction that could reduce the need for laboratory testing, but the evidence mainly shows improved average accuracy on synthetic networks under the assumed error model. The real-network comparison in §5.4 is qualitative: on CRN-T the averaged solver assigns about 23% of the mass to the observed product, but the experimental yield is not reported, so the size of the stated underestimate cannot be judged. Reporting the predicted and experimental yields side by side, or framing the contribution more modestly, would make the claims and the evidence match.
  • Report the distribution of predicted populations and evaluate its calibration, not only its mean.
2 comments
AuthorsOpenReview ↗

Response to Reviewer A2a2 1/2

We thank the reviewer for their careful and constructive feedback, we address each point raised by the reviewer below.

Testing the gains under an error model different from the one used to build the method

The improvements are demonstrated mainly under the same error model used to build the method. Not when the error model is wrong.

We thank the reviewer for this point. Fig. 2b partially addressed it by varying the assumed σ′ for a fixed true σ. We have extended this into a sweep over σ x σ′ (Appendix E.1): p̄ improves over p for all 0 < σ′ ≤ 2σ, for σ ∈ {3, 6, 12} kcal/mol. p̄ therefore needs only a rough estimate of the barrier uncertainty, not a calibrated one. To test errors outside p̄'s noise model, we also added an experiment with systematic barrier bias, described next.

Averaging over perturbations around a biased estimate does not correct the bias, and whether it still improves the predicted populations in that case is not tested.

The proposed uncertainty aware model is not designed to correct bias. It targets the random part of the barrier error. Bias can often be reduced separately by calibration against reference data. However, we agree that it is important to check whether the model still helps when bias is present. We therefore added an experiment (App. E.2) including a systematic barrier bias b that p̄ does not account for. For b ∈ [−12, 12] kcal/mol and σ ∈ {3, 6, 12} kcal/mol, p̄ still achieves 8-24% lower population RMSE than p.

CRN-T validation

On CRN-T, the averaged solver assigns some population to the observed product, but this does not show that the predicted populations are quantitatively accurate

We agree that this result shows qualitative rather than quantitative agreement. The experimental rate constant (k > 10^6 M^-1 s^-1; Dowideit & von Sonntag, 1998) implies that at least 57% of the reactant converts to product within 1 ms. The deterministic solver p predicts 0.5% conversion, while the proposed q̄ predicts 23%, which is substantially closer to experiment. We have added this to Section 5.4. To our knowledge, no large public dataset of reaction networks with measured populations exists, so our quantitative evaluation is based on thousands of synthetic networks (Sections 5.1-5.2) and CRN-T serves only as a qualitative check.

the assumed uncertainty (σ′ = 4 kcal/mol) is chosen rather than calibrated.

We agree. To check that our conclusions do not depend on this choice, we repeated the analysis for σ′ ∈ {2, 4, 6, 8} kcal/mol (App. E.3). The improvement over p persists across this range. We did not calibrate σ′ because this would require experimental data for many reactions, while the source study (Petrus et al., 2025) provides only two.

Gains at low noise

The gains are also modest at lower noise levels (Fig. 5)

This is expected by design. q̄ accounts for errors in the input barriers, so when the noise is low there is little error to account for, and both solvers converge to the same result. At noise levels typical of current electronic structure methods (σ ≈ 3-14 kcal/mol), q̄ reduces population RMSE by 13–21% (Fig. 4, M = 100).

Reference baselines

e.g. retaining the initial reactant or using equilibrium populations

We thank the reviewer for this suggestion. We will report these baselines in the camera ready version of the paper.

Speed up and stopping rule

…it is not compared with standard tolerance-based Krylov stopping at matched accuracy.

We thank the reviewer for this point. q̄ and p̄ match the same population RMSE in all experiments in Sections 5.1-5.3. This was stated only briefly at the start of Section 5. We have now quantified this in Section 5.1: their population RMSEs differ by at most 1e-4 at every M and σ. The SNR cutoff therefore adds no measurable error, and the speed-up comes at matched accuracy.

computing the barriers themselves takes far longer (Table 1). The paper does not show a workflow in which this saving matters in practice.

Our main contribution is the accuracy gain: q̄ gives, in ~100 ms of solver time, the RMSE gain that would otherwise take days of extra electronic structure computation. Running q̄ on barriers from a cheaper method (10.4 kcal/mol error, 8 h per step) reaches the same population RMSE as running p on barriers from a more accurate method (6.5 kcal/mol error, 1.5 days per step) (Fig. 4, Table 1). As the reviewer notes, solver time is small compared with computing the barriers, which we see as a strength: this gain comes at negligible extra cost. The stopping rule is an additional contribution that keeps this cost low. We have revised the timing discussion in Section 5.3 to make this explicit.

AuthorsOpenReview ↗

Response to Reviewer A2a2 2/2

Summary remarks:

We thank the reviewer for their constructive feedback. To address the main concern, how the method behaves under a misspecified error model, we added three experiments: p̄ improves over p for all 0 < σ′ ≤ 2σ (App. E.1); it reduces population RMSE by 8–24% under an unmodelled systematic bias b ∈ [−12, 12] kcal/mol (App. E.2); and on CRN-T the improvement holds for σ′ ∈ {2, 4, 6, 8} kcal/mol (App. E.3). We also show that the stopping rule adds no measurable error (Section 5.1). In practice, q̄ can be used as a drop-in replacement for standard kinetic solvers: practitioners only need to add a rough estimate of the barrier uncertainty to get more accurate estimates at negligible extra cost.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewer

We thank all reviewers for their feedback. Reviewers found the paper well motivated, and the empirical evaluation systematic. The main concerns were shared across reviewers, we have addressed them as follows:

  • Robustness to a misspecified error model: We added a σ × σ' sweep and an experiment in which the barrier errors include a systematic bias that p̄ doesn't model. In the reported experiment p̄ improves over p wherever 0 < σ' ≤ 2*σ (Appendix E.1), so a rough estimate of the uncertainty is enough to benefit. And under bias it still achieves lower population RMSE than p (Appendix E.2).

  • CRN-T quantitative yields and choice of σ′: We now report the quantitative yields and sweep σ' ∈ {2, 4, 6, 8} kcal/mol. The improvement of q̄ over p persists across the range (Appendix E.3).

  • Accuracy of the SNR truncation: We did compare q̄ with p̄, in every experiment 5.1-5.3, but stated this only briefly and was overlooked. We now report it explicitly in Section 5.1: With the same networks and barrier samples, their population RMSEs differ by at most 1e-4. The 1.3 to 2.3× speed-up therefore comes at no measurable loss of accuracy.

  • Role of Appendix B and SNR truncation (Eq. 10): Eq. 10 is an empirical stopping rule computed directly from the sampled populations and does not depend on Appendix B. Appendix B only supports an equation that builds intuition in Section 4.2. We believe our original text did not make this clear which created some confusion. We have stated it now clearly in Section 4.2.

  • Relation to Gaussian Smoothing literature: We now cite additional missing papers from the Gaussian smoothing literature. Our Related Work (Section 6) already discusses prior uncertainty propagation in reaction networks. Those works use the uncertainty for error bars and refinement, while we use it in the prediction itself.

We hope these revisions address the reviewers' main concerns. Our main contribution is a kinetic solver that leverages barrier uncertainty to make more accurate prediction. In practice, q̄ can be used as a drop-in replacement for standard kinetic solvers, practitioners only need to add a rough estimate of the barrier uncertainty to get more accurate estimates at negligible extra cost.

As reaction networks become increasingly generated automatically and at scale (Section 6), we hope q̄ offers a practical, low-cost improvement for practitioners working on computed reaction kinetics.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.