{"id":"7f03437d-641d-4f26-aedc-3bc968702e1d","arxiv_id":"2607.16129","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Quantum Monte Carlo-trained neural-network force fields enable multi-state nonadiabatic dynamics of azomethane, reducing CASSCF's overestimated C–N dissociation and predicting a small prompt dissociation component.","lead":"This paper trains machine-learned force fields on quantum Monte Carlo (QMC) data to run photochemical dynamics of azomethane without the cost of direct QMC trajectories. The QMC-based model suppresses the excessive C–N bond breaking seen with CASSCF and predicts a small early dissociation component after internal conversion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QMC-ML trans-azomethane dynamics extrapolates a model trained only on cis-initiated CASSCF geometries; the reported 9.5% yield and ~160 fs onset may be extrapolation artifacts.","rationale":"Good-faith reading: This is a well-executed study. The VMC/CIPSI reference is benchmarked against exFCI for vertical excitations; force benchmarks along isomerization and dissociation paths are informative; 1000-trajectory ensembles with error bars add statistical rigor; the MS-CASPT2 ML model is checked against ab initio dynamics. The central methodological advance—using QMC references to train multi-state ML force fields—is credible.\n\nThe most load-bearing soft spot is not the QMC method itself but the domain where the trained model is applied. The trans-azomethane dynamics, which provides the direct comparison to femtosecond experiments, is performed with a model trained exclusively on geometries from cis-initiated CASSCF dynamics. The authors' justification is that the two isomers share the same torsional coordinate and dissociation channels. This is reasonable but untested. If the trans trajectories populate regions outside the training distribution, the dissociation yield and onset time—quantitative outputs central to the paper's abstract—would be extrapolations. Because the QMC ML model cannot be directly validated against on-the-fly QMC dynamics (the very impracticality that motivates the ML surrogate), the training coverage is the main vector by which the QMC reference could fail to affect the dynamics.\n\nThe reader's weakest_assumption framed this as a general CASSCF-generated coverage bias. I agree, and the trans application is the sharpest manifestation because it involves a different isomer and a direct experimental comparison. A retraining experiment with trans-specific adaptive sampling would settle whether the reported 9.5% / ~160 fs are robust. Without that, a conditional verdict is appropriate.\n\nI do not think this rises to rejection: the cis dynamics and the core methodological demonstration survive even if the trans numbers shift. The verdict should remain CONDITIONAL with the added requirement of trans-specific training or a clear demonstration that trans-visited geometries lie inside the training distribution.","tokens_in":17670,"tokens_out":8028,"duration_ms":72436,"concrete_test":"Run a trans-specific active-learning loop: start from trans-azomethane Wigner geometries, propagate 100–200 trajectories with the current QMC-ML model, collect the visited geometries (especially around the ~160 fs onset), compute VMC/CIPSI energies/forces for them, add these to the training set, retrain the SpaiNN model, and rerun the 1000-trajectory trans dynamics. If the dissociation yield (9.5%) or the onset (~160 fs) changes by more than the statistical uncertainty (≈±2% yield; ±20 fs onset), the original trans result is an extrapolation artifact; if not, the coverage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III D states the models are 'trained on the dataset generated from the cis-initiated dynamics' and used for trans 'without additional trans-specific training.' The justification given—that both isomers probe the same torsional coordinate and C–N cleavage channels—is plausible but not demonstrated. Trans starts from a different equilibrium geometry, and its Wigner-sampled initial conditions plus post-IC trajectories may visit regions (e.g., the trans-side basin, different C–N stretch momenta) that are sparse or absent in the cis-generated 2320-configuration dataset. The key experimental comparison (Diau–Zewail, 70–100 fs onset) is made against this trans simulation (Fig. 8, onset ~160 fs). If the model must extrapolate in the trans region, the 9.5% dissociation yield and the ~160 fs onset—the main evidence for a 'prompt dissociation component'—could be artifacts of training-set coverage rather than QMC physics. The paper's own Section V limits the claim to 'qualitative agreement,' but the abstract's 'timescale consistent with experiments' is stronger. This is a concrete instance of the broader CASSCF-generated coverage bias: the training geometries were generated once by CASSCF adaptive sampling, so any region where QMC dynamics differs from CASSCF is potentially under-sampled.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a workflow in which variational Monte Carlo (VMC) wave functions built from CIPSI expansions are used as reference data to train multi-state neural-network force fields for nonadiabatic molecular dynamics. The method is applied to gas-phase azomethane, with the QMC-based model compared against CASSCF- and MS-CASPT2-based models. The authors benchmark vertical excitation energies and forces along isomerization and dissociation pathways, then propagate 1000-trajectory surface-hopping ensembles from both cis- and trans-azomethane. Their central findings are that the QMC-trained dynamics preserves the torsional relaxation through conical intersections, strongly reduces the excessive C–N dissociation seen with CASSCF (from 92% to 33% at 400 fs for cis, and 82% to 9.5% for trans), and predicts a small prompt dissociation component after internal conversion with an onset around 160 fs, in qualitative agreement with femtosecond experiments.","tokens_in":18022,"tokens_out":3287,"duration_ms":32550,"significance":"If the results are robust, the paper establishes a practical route for large-ensemble photodynamics with correlated wave-function reference data, which is a significant methodological advance. The force benchmarks are careful and informative: they demonstrate that fixed-active-space methods have geometry-dependent errors near conical intersections and along dissociation, whereas the VMC/CIPSI protocol with a fixed target PT2 correction behaves more consistently. The large trajectory ensembles (1000 per method) are a strength, as is the explicit comparison with exFCI for vertical excitation energies. The main risk concerns extrapolation of the ML model outside the configuration space spanned by the CASSCF-generated training set, which is directly relevant to the headline dissociation yields and the claimed prompt-dissociation timescale.","major_comments":[{"comment":"The trans dynamics is performed with an ML model trained exclusively on the dataset generated from cis-initiated CASSCF adaptive sampling, as stated in Section IV.D. The justification that cis and trans trajectories probe the same torsional coordinate and C–N cleavage channels is plausible but not demonstrated. The trans simulation provides the key experimental comparison (Fig. 8 onset ~160 fs vs Diau–Zewail 70–100 fs), and the 9.5% dissociation yield and the claimed prompt component rest entirely on this model. If the model must extrapolate in the trans-side basin or in post-internal-conversion regions that are sparse in the 2320-configuration cis-generated set, these results could be artifacts of training-set coverage. Please provide quantitative evidence: for example, project the training set and the trans-initiated trajectories onto the ∠CNNC dihedral and C–N bond-length coordinates,","section":"Section IV.D (trans-azomethane dynamics) and Section III (adaptive sampling)"},{"comment":"The 2320 configurations used to train all three ML models are generated once by adaptive sampling at the CASSCF(12,10) level, and the QMC, CASSCF, and CASPT2 reference data are computed on the same geometries. This creates a coverage bias: any region of configuration space that the QMC potential-energy surfaces visit but CASSCF dynamics does not will be under-sampled, and the QMC-trained force field must extrapolate there. Since the central claim is that QMC substantially changes the dynamics relative to CASSCF, the very regions where the method is most important may be exactly the ones not sampled. I ask the authors to demonstrate that the QMC-ML trajectories stay within the training distribution (e.g., by monitoring a distance-to-training-set metric or comparing the distribution of key internal coordinates visited in the QMC dynamics with those in the training set), or to augment the s","section":"Section III (training set generation, final paragraph)"},{"comment":"For dissociated geometries, the excited-state force deviations of CASSCF and MS-CASPT2 relative to VMC/CIPSI exceed 20 kcal/mol/Å. The authors argue that this should not directly affect the dissociation dynamics because C–N breaking occurs on the ground state. However, the ML models are multi-state and the state-averaged orbitals from the excited-state description enter the ground-state potential through the common determinantal space and the training of the coupled ML model. Given that dissociation yields are a central quantitative result, this claim should be supported quantitatively — for example, by showing how the excited-state force errors propagate into ground-state forces at the sampled dissociation geometries, or by comparing dynamics with and without the problematic excited-state training points.","section":"Section IV.B, Fig. 3 and surrounding discussion"}],"minor_comments":[{"comment":"The abstract states that the prompt-dissociation timescale is \"consistent with\" experiments, while Section V concludes \"qualitative agreement\" and notes dependence on the bond-length criterion and initial sampling. Please harmonize these phrasings so the abstract does not overstate the strength of the comparison.","section":"Abstract / Section V"},{"comment":"Minor typo: \"neural-network force fields [25] as a smooth surrogates\" should read \"as smooth surrogates.\"","section":"Section II.A / intro"},{"comment":"The caption contains \"cis-iniziated\" — should be \"cis-initiated.\"","section":"Figure 6 caption"},{"comment":"For the trans rows, QMC yields are reported without uncertainty (e.g., 9.5% of 1000 trajectories has a 95% binomial confidence interval of roughly ±1.8%). Reporting the uncertainty would help assess the significance of differences between methods.","section":"Table II"},{"comment":"The exFCI values are used as benchmark, but they are produced within the same CIPSI family. Although the paper is transparent about this, a sentence noting that exFCI is an extrapolated limit from the same selected-CI framework would help the reader judge the degree of independence.","section":"Section IV.A and Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong methodological contribution with careful benchmarks, but the headline quantitative claims (dissociation yields and the prompt onset) depend on extrapolation of an ML model trained on CASSCF-generated geometries. The trans result in particular is the linchpin of the experimental comparison and is the least supported by data coverage. If the authors can convincingly demonstrate that the QMC-ML trajectories remain within the training distribution, or provide trans-specific training and show robustness, the paper would be a solid addition. I recommend major revision rather than rejection because the central approach is sound and the concern is addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core method is genuinely new: a multi-state machine-learned force field trained on VMC/CIPSI energies and forces, used for surface-hopping dynamics. That is a real step beyond ground-state QMC-ML, and azomethane is a demanding test with conical intersections and bond breaking. Second, the trans-azomethane results are softer than the abstract implies, for a concrete reason the authors mostly acknowledge but do not resolve.\n\nThe paper does a lot well. The vertical excitation energies match exFCI to within 0.05 eV, and the force benchmarks along isomerization and dissociation paths are thoughtful. The dynamics comparison uses 1000 trajectories per method and shows large, systematic differences: CASSCF massively over-dissociates, CASPT2 is intermediate, and the QMC-trained model suppresses dissociation while preserving the torsional mechanism. The QMC data are ab initio, not fit to experiment, so the dynamics is a genuine prediction. That is a meaningful contribution.\n\nThe main soft spot is the trans simulation. The models are trained on configurations generated by CASSCF adaptive sampling from cis-initiated dynamics, then applied to trans without any trans-specific training. The authors argue both isomers probe the same torsional coordinate and dissociation channels, which is plausible but not demonstrated. The experimental comparison for prompt dissociation uses the trans runs, and those results are exactly the ones that could be affected by training-set coverage. The 9.5% yield and ~160 fs onset may be real QMC physics, or they may be artifacts of extrapolation. The paper's own conclusion says \"qualitative agreement,\" while the abstract says \"consistent with femtosecond-resolved mass-spectrometry experiments\"—the former is accurate, the latter is too strong.\n\nTwo smaller issues: the dataset, hyperparameters, and trained models are not deposited, which makes independent reproduction hard; and the PT2-based CIPSI target is the same for all geometries, which is sensible but should be more carefully defended as a universal accuracy control.\n\nOverall, this is a solid, honest methods paper with an important result. It deserves serious peer review. I would ask for data deposition, a softened abstract, and either retraining or explicit validation on trans-sampled geometries before accepting. The cis story stands on its own even if the trans question remains open.","headline":"First credible multi-state QMC-trained ML force fields for nonadiabatic dynamics; the cis-azomethane story holds up, but the trans results lean on an extrapolated model and the abstract overstates the experimental match.","tokens_in":18512,"tokens_out":1527,"would_cite":true,"duration_ms":15633,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum Monte Carlo-trained neural force fields bring correlated wave-function accuracy to photodynamics: for azomethane they cut predicted C-N dissociation from 92% to 33% (cis) and 82% to 9.5% (trans), with a ~160 fs prompt dissociation c","keywords":["quantum Monte Carlo","machine-learned force fields","nonadiabatic dynamics","surface hopping","azomethane photodynamics","conical intersections","C-N bond dissociation"],"falsifier":"Run direct on-the-fly VMC surface-hopping trajectories, without the neural-network surrogate, for a few dozen cis-azomethane initial conditions through the first conical intersection and out to roughly 400 fs; if the fraction of trajectories with a C-N bond longer than 2.25 Å, or the onset time of first bond cleavage, differs substantially from the ML model's ~33% and ~160 fs, the training-coverage assumption is falsified.","tokens_in":17496,"feed_emoji":"⚛️","tokens_out":9533,"duration_ms":81700,"temperature":0.7,"pith_summary":"The paper aims to establish that variational quantum Monte Carlo (QMC) can serve as a practical reference for machine-learned nonadiabatic photodynamics: neural-network force fields trained on QMC energies and forces can propagate large ensembles of surface-hopping trajectories with correlated wave-function accuracy. For azomethane, this changes the chemical story: fixed-active-space CASSCF predicts runaway C-N bond breaking (92% of cis trajectories within 400 fs), while the QMC-trained model keeps the torsional photoisomerization mechanism central and reduces dissociation to 33%; for trans-azomethane the reduction is even larger, from 82% to 9.5%. The same dynamics yields a small prompt dissociation component after internal conversion, with C-N cleavage beginning around 160 fs, in qualitative agreement with femtosecond-resolved mass-spectrometry experiments. If correct, this makes QMC-trained machine learning a practical route to correlated excited-state dynamics without the manual active-space choices that dominate current errors.","feed_headline":"QMC-trained dynamics cuts azomethane bond breaking from 92% to 33%","feed_subtitle":"For the trans isomer the yield drops to 9.5%, and early C-N cleavage appears near 160 fs.","key_machinery":"The load-bearing object is a Jastrow-Slater variational wave function whose determinantal part comes from selected configuration-interaction expansions, optimized state-specifically with an orthogonality penalty and targeted at a common second-order perturbation correction across all geometries and states. The stochastic VMC energies and forces are converted by a neural-network force field into smooth multi-state potential-energy surfaces, which are then used in trajectory surface-hopping dynamics. Crucially, the determinantal expansion adapts its size to the geometry—hundreds of determinants near the cis minimum, only about 150 for dissociated fragments—so the method keeps balanced accuracy","core_discovery":"The paper's central claim is that stochastic variational Monte Carlo energies and forces, obtained from Jastrow-Slater wave functions with selected configuration-interaction expansions, can be turned by neural networks into smooth multi-state force fields and used for large-ensemble surface-hopping photodynamics. On azomethane, this changes the computed product distribution: CASSCF puts 92% (cis) and 82% (trans) of trajectories on a C-N dissociative path within 400 fs, while the QMC-trained model yields 33% and 9.5%, respectively, while preserving the expected torsion to the conical-intersection region. The QMC dynamics also shows a small but non-negligible prompt dissociation component afte","pith_inferences":["Inference: The stated yields may carry a training-coverage bias, because all QMC labels were computed at geometries that CASSCF adaptive sampling selected; a self-consistent extension would let the QMC-trained model propose new geometries and add QMC labels there, especially along the dissociation asymptote.","Inference: The ~160 fs prompt component is sensitive to the C-N bond-length criterion (2.25 Å here) and to Wigner sampling with zero-point energy; recomputing yields with a shorter threshold and classical Boltzmann sampling would show how much of this component is physical versus sampling-driven.","Inference: The same pipeline should transfer to other photochemical systems where active-space methods degrade at conical intersections or along bond-breaking coordinates, making method-dependent dynamics the norm rather than the exception."],"forward_implications":["CASSCF's excess dissociation is a systematic error of the fixed-active-space reference rather than a sampling artifact, since the QMC and CASPT2 models share the same geometries and initial conditions yet fragment much less.","The torsional mechanism (normal and rotator pathways through two symmetry-related conical intersections) is robust across electronic-structure methods, shifting the open question to quantitative branching and dissociation yields.","The predicted ~160 fs onset of C-N cleavage after internal conversion gives time-resolved experiments a specific target to confirm or rule out an impulsive dissociation component.","A QMC-trained ML force field can be applied to the trans isomer without retraining, covering both photoisomerization directions with one correlated reference dataset.","One thousand-trajectory, 400 fs surface-hopping ensembles with correlated reference data are computationally feasible through the ML surrogate."],"fun_headline_variants":["QMC-ML slashes azomethane C-N breaking from 92% to 33%","Quantum Monte Carlo-trained ML fixes azomethane bond-breaking","Azomethane test: QMC-trained dynamics cut C-N cleavage to 33%","QMC-ML predicts 9.5% trans C-N yield in azomethane"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire QMC training set is generated once from CASSCF adaptive-sampling trajectories, so if the true correlated surface visits geometries CASSCF never explores (for example different torsional or dissociation pathways), the QMC-trained force field must extrapolate beyond its training data, and the dissociation yields plus the ~160 fs prompt onset inherit that coverage bias.","fun_headline_variants_meta":{"raw":{"variants":["QMC-ML slashes azomethane C-N breaking from 92% to 33%","Quantum Monte Carlo-trained ML fixes azomethane bond-breaking","Azomethane test: QMC-trained dynamics cut C-N cleavage to 33%","QMC-ML predicts 9.5% trans C-N yield in azomethane"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001095,"raw_usage":{"total_tokens":4422,"prompt_tokens":770,"completion_tokens":3652,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":3575}},"tokens_in":514,"tokens_out":3652,"duration_ms":25531,"temperature":1.0,"reasoning_tokens":3575,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:15:18.714910+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run direct on-the-fly VMC surface-hopping trajectories, without the neural-network surrogate, for a few dozen cis-azomethane initial conditions through the first conical intersection and out to roughly 400 fs; if the fraction of trajectories with a C-N bond longer than 2.25 Å, or the onset time of first bond cleavage, differs substantially from the ML model's ~33% and ~160 fs, the training-coverage assumption is falsified.","supporting_citations":[],"review_version":1}