{"id":"ec14457b-ee5b-4fe6-8da4-6b45e0c883f9","arxiv_id":"2411.10075","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Thermodynamic Interpolation learns temperature-conditioned flows between Boltzmann ensembles, allowing sampling, free energy differences, and gEDMD kinetic rates at held-out temperatures.","lead":"This paper presents Thermodynamic Interpolation, a generative method that learns maps between molecular ensembles at different temperatures using normalizing flows, enabling sampling, free energy estimates, and kinetic rates at new temperatures. It is a candidate tool for computing thermodynamic and kinetic properties without long low-temperature simulations, but a central equation in the method description appears to have a sign error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (20) as printed is internally inconsistent: at equal temperatures with the identity map it gives weights exp(-2E/kT), not 1, so the aTI unbiased-reweighting construction is not reproducible as written.","rationale":"The reader's rejection is correct, but the strongest reason is not the overdamped-Langevin modeling assumption highlighted as the weakest assumption. That limitation is explicitly acknowledged in the text ('assuming the generator is Brownian') and the kinetics section is framed as approximate. The truly load-bearing defect is Eq. (20): it is the mathematical core of aTI's unbiasedness, and it fails a trivial identity check. The correct importance weight has +E(xA)/(kTA); the printed weight has -E(xA)/(kTA). This cannot be dismissed as a choice of convention: at equal temperatures with the identity map it produces weights proportional to exp(-2E/kT), which reweight the source ensemble to low energies rather than leaving it unchanged. If the authors' implementation uses the corrected formula, then the manuscript needs a corrigendum before the method can be reproduced; if it uses the printed formula, the reported ESS and observables are biased. The empirical results do not resolve this because no code is provided. I also note the reader's rationale already flags this sign error, so I agree with the REJECT verdict; the weakest_assumption field points to a different, less central issue, hence partial agreement.","tokens_in":15454,"tokens_out":8327,"duration_ms":92505,"concrete_test":"Independently re-derive Eq. (20) and evaluate the identity limit: for a trained ADW aTI map, take any sample xA, set TA=TB and f_AB=id, and compute the printed weight; if it differs from 1 by more than numerical noise, the estimator is biased as stated. Then rerun the N-Me ESS and Delta F results of Figs. 4B-C using the corrected exponent +E(xA)/(kTA) - E(f_AB(xA))/(kTB) + log|det J|, without IQR filtering; if the conclusions change materially, the empirical success rests on an equation not present in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step for the ambient-TI sampling claim is the importance weight in Eq. (20): wa(xA,TA,TB)=exp(-E(xA)/(kTA) - E(f_AB(xA))/(kTB) + log|det J|). A change-of-variables derivation from xA~mu_A to y=f_AB(xA) gives an unnormalized target/source density ratio proportional to exp(-E(y)/(kTB) + E(xA)/(kTA) + log|det J|), so the first energy term must be positive. The printed expression has both energy terms negative. The identity limit is decisive: set TA=TB and f_AB=id; the correct weight is 1 up to the usual partition-function constant, while Eq. (20) gives exp(-2E(x)/(kT)), which after normalization preferentially weights low-energy conformations and is not the target Boltzmann measure. Since Eq. (20) is the basis for the reported ESS values and for unbiased observables from aTI, the central claim of unbiased temperature-interpolated sampling is not supported by the manuscript as written. This is an internal inconsistency, not a disagreement with external consensus. The same issue is noted in the reader's rationale; the reader's stated weakest assumption (overdamped Langevin generator) is a separate and explicitly caveated approximation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces Thermodynamic Interpolation (TI), a generative method for sampling molecular configurations across thermodynamic states. Two variants are presented: ambient TI, which learns a direct diffeomorphic map between Boltzmann ensembles at different temperatures in configuration space, and latent TI, which maps from a normal latent distribution to a temperature-conditioned ensemble. The velocity fields are trained with stochastic interpolants and SE(3)-equivariant message passing networks. The authors validate TI on a 1D asymmetric double well and two small molecules (N-methylformamide and 3-propan-2-ylhex-1-yne), reporting high effective sample sizes, accurate free energy differences via targeted free energy perturbation, and temperature-dependent kinetic rates obtained from generator EDMD. The paper claims interpolation and extrapolation in temperature, including temperatures not seen during training.","tokens_in":15688,"tokens_out":5392,"duration_ms":53601,"significance":"If fully correct, the TI framework would be a useful addition to the Boltzmann generator toolbox: it shares information across temperatures, avoids replica ladders, and connects generative sampling to TFEP and gEDMD. The paper's strengths are its grounding in the stochastic interpolant framework, the use of standard TFEP relations, and held-out temperature validation against independent MD simulations and numerical integration. Experimental comparisons are plausible, and the authors are transparent about some limitations (Jacobian cost, Brownian generator assumption). However, the central importance-weight formula for ambient TI contains a sign error that invalidates the unbiasedness argument as written, so the quantitative claims must be re-evaluated. The kinetics section is also explicitly model-dependent and should be framed as Brownian dynamics rates rather than true molecular rates.","major_comments":[{"comment":"The importance weight in Eq. (20) is printed with two negative energy terms. A change-of-variables derivation from xA ~ mu_A to y = f_AB(xA) gives a target/source density ratio proportional to exp(-E(y)/(kTB) + E(xA)/(kTA)) times |det J|, so the first energy term must be positive. In the identity limit TA=TB and f_AB=id, the correct weight is 1 up to the usual partition-function constant, whereas Eq. (20) yields exp(-2E(x)/(kT)), which after normalization preferentially weights low-energy conformations and is not the target Boltzmann measure. Since Eq. (20) is the basis for the reported aTI ESS values and for unbiased observable estimates from ambient TI, the central sampling claim is not supported by the manuscript as written. The formula must be corrected and the numerical results re-examined.","section":"Methods, Ambient TI, Eq. (20)"},{"comment":"The kinetic-rate estimates are derived under the overdamped Langevin generator in Eqs. (24)-(25), and the text states \"assuming the generator is Brownian.\" For the molecular systems, the reference MD is underdamped, so the estimated rates are not true molecular transition rates; the paper's own MSM comparison for the ADW system shows that overdamped rates bound underdamped rates from below. The claims about \"kinetic exchange rates\" for molecules should be explicitly restricted to the Brownian model, and the comparison to MD-derived rates obtained with the same misspecified estimator does not validate true kinetics. This is a caveated but load-bearing limitation for the kinetics section.","section":"Methods, gEDMD; Results, Generator Estimation"},{"comment":"The authors state that they use an IQR-filtering strategy to \"avoid high variance on the free energy estimators\" (Figures 3B-E and 4B-C). If outliers are removed based on the values of the estimator or the importance weights, this can bias the reported free energies and effective sample sizes. The Supporting Information must specify the filtering rule, clarify whether it depends on the estimator, and demonstrate (e.g., by reporting unfiltered or alternatively robust estimates) that the reported results are not an artifact of the filter.","section":"Results, Figures 3 and 4; IQR filtering"}],"minor_comments":[{"comment":"The term \"Thermodynamic Interpolation\" is introduced twice, and the sentence \"We introduce TI... We introduce TI flavors\" is redundant; please tighten the wording.","section":"Abstract and Introduction"},{"comment":"The molecule is referred to as both \"3p2y1y\" (abstract and Figure 5) and \"3p2y1p\" in the main text; please make the abbreviation consistent.","section":"Results, first paragraph"},{"comment":"The normalized temperature \\hat T is defined with an overline T that is not explicitly defined; please define all symbols in the equation.","section":"Methods, Eq. (21)"},{"comment":"The statement that exact computation of free energies \"is not tractable in the current architectures\" for larger systems is a strong limitation; please quantify the scaling for the tested systems or add a reference to the stated linear Jacobian cost.","section":"Conclusion"},{"comment":"The \"TOC Graphic\" appears twice on the final pages, which appears to be a duplication in the manuscript.","section":"Manuscript formatting"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the incorrect importance-weight formula in Eq. (20). Because the error is local and appears correctable, I recommend major revision rather than rejection, but the numerical results must be recomputed with the corrected weight before the central claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper deserves a serious referee, but not as it stands. The core idea—temperature-conditioned stochastic interpolants that map directly between Boltzmann ensembles in ambient space—is genuinely new and worth building on. The held-out temperature experiments are the right kind of evidence: the models are trained on a range of temperatures and then predict a temperature left out, comparing against independent MD simulations. The double-well results show high ESS and accurate free energies with far fewer samples than direct reweighting, and the molecular torsion distributions match reference histograms well. The authors also cite related work fairly and acknowledge the main assumptions in the kinetics part.\n\nNow the soft spots. The load-bearing problem is Eq. (20), the ambient-TI importance weight. As printed, it is exp(-E(x_A)/kT_A - E(f_AB(x_A))/kT_B + log|det J|). Both energy terms are negative. A simple change-of-variables calculation gives the correct unnormalized target/source ratio as exp(-E(f(x))/kT_B + E(x)/kT_A + log|det J|)—the first energy term must be positive. The identity limit is decisive: set T_A = T_B and f = id; the correct weight is 1 (up to a normalization constant), while Eq. (20) gives exp(-2E/kT), which preferentially downweights high-energy configurations and is not the Boltzmann measure. Since this equation is the basis for the reported ESS values and for unbiased observables from aTI, the central claim is not reproducible from the manuscript as written.\n\nOther issues are smaller but still worth raising. The abstract and conclusion claim extrapolation beyond the training-temperature range, but every held-out temperature in the experiments lies inside the training interval. That is interpolation, not extrapolation. The IQR filtering of free-energy estimates is a post-hoc step that can bias the estimator if not handled carefully; the paper should justify it more rigorously. And the kinetic rates depend on the overdamped Langevin generator, an approximation the authors do flag, but the approximation is not validated against underdamped reference rates in the main text.\n\nWho is this for? Computational chemists and physicist working on generative models for molecular ensembles. The idea is plausible and the experimental setup is honest, so the sign error looks like a fixable typo rather than a fundamental flaw. I would send it to peer review with the expectation that the authors correct Eq. (20), release code, and temper the extrapolation claims. Desk rejection would be premature.","headline":"A promising temperature-interpolating flow method, but the printed importance weight in Eq. (20) has both energy terms negative, which breaks the unbiased sampling construction as written.","tokens_in":16254,"tokens_out":2420,"would_cite":false,"duration_ms":27653,"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":"Thermodynamic Interpolation learns a differentiable map between Boltzmann ensembles at different temperatures, and the paper claims this map generates accurate ensembles, free energies, and approximate kinetic rates at temperatures not…","keywords":["Thermodynamic Interpolation","normalizing flows","Boltzmann generators","free energy perturbation","gEDMD","kinetics","temperature generalization","stochastic interpolants"],"falsifier":"Run long unbiased underdamped molecular dynamics for N-methylformamide at 300 K, build a Markov state model, and compare its slowest relaxation rate with the gEDMD rate obtained from TI-generated samples at the same temperature; a substantial mismatch would localize the failure in the Brownian-generator assumption rather than the learned map.","tokens_in":15214,"feed_emoji":"🔥","tokens_out":6510,"duration_ms":60177,"temperature":0.7,"pith_summary":"Thermodynamic Interpolation (TI) is a generative method for mapping samples between Boltzmann distributions at different temperatures. The paper claims that a learned diffeomorphism, trained on molecular configurations at a few temperatures, transforms high-temperature ensembles into accurate low-temperature ensembles at temperatures never seen in training, with high effective sample size. It further claims that these transformed samples yield accurate free-energy differences via targeted free energy perturbation, and that combining TI with generator extended dynamic mode decomposition (gEDMD) gives approximate kinetic rates whose temperature dependence matches reference simulations. The goal is to make one learned model stand in for many costly low-temperature simulations while giving both thermodynamics and kinetics across a temperature range.","feed_headline":"Flow maps between temperatures yield molecular thermodynamics","feed_subtitle":"One learned map predicts free energies and kinetic rates at temperatures never simulated.","key_machinery":"The central object is the learned diffeomorphic map $f_{AB}$ between Boltzmann distributions $\\mu_A$ and $\\mu_B$ at temperatures $T_A$ and $T_B$, implemented as a continuous normalizing flow with an SE(3)-equivariant ChiroPaiNN velocity field $b^{(\\theta)}(t,x;T_A,T_B)$ trained by the stochastic interpolant objective. The two-sided interpolant $x(t) = (1-t)x_A + t x_B + \\gamma(t) z$ defines a stochastic path between paired samples, and minimizing the regression loss (eq. 19) yields a velocity whose integrated ODE transports $x_A \\sim \\mu_A$ to $x_B \\sim \\rho_B$. The same flow supplies the log-Jacobian determinant entering the importance weights $w_a$ (eq. 20), which correct the surrogate samples to unbiased estimates at the target temperature. For kinetics, the machinery is the gEDMD estimator of the Koopman generator for overdamped Langevin dynamics, whose eigenvalues give relaxation rates, applied to the TI-generated equilibrium samples.","core_discovery":"The central claim is that a continuous normalizing flow can learn a velocity field, trained by stochastic interpolants, that maps a Boltzmann ensemble at one temperature directly onto another, and that this map generalizes to target temperatures outside the training set. On the model systems studied (a one-dimensional asymmetric double well and two molecules from the QM9 set, N-methylformamide and 3-propan-2-ylhex-1-yne), the ambient TI map produces torsion-angle histograms and TICA projections that match reference low-temperature MD, and Kish effective sample sizes remain high enough that free-energy differences from as few as five transformed samples match what direct reweighting needs roughly twenty times more samples to achieve. The latent TI variant, a temperature-conditioned Boltzmann generator, samples multiple temperatures on demand from a shared latent Gaussian. The authors also show that gEDMD, applied to TI-generated equilibrium samples, yields rates for slow metastable exchanges whose temperature trends agree with rates estimated from MD, provided the underlying dynamics are taken to be overdamped Langevin (Brownian) dynamics.","pith_inferences":["The same learned-map architecture should extend to thermodynamic variables other than temperature, such as pressure or chemical potential, since the flow is conditioned on the state parameters; this is a natural but untested extrapolation of the paper's approach.","Because the paper finds gEDMD rates are robust to mapping bias even without reweighting, kinetic studies might skip importance weights entirely, accepting a small systematic error for large computational savings; the paper hints at this but does not recommend it.","The high sample efficiency at very small sample counts suggests TI could be applied to expensive potential-energy surfaces, such as quantum-mechanical calculations, where generating many MD frames is prohibitive.","Multi-temperature training may reduce the amount of simulation data needed per temperature, mirroring the information sharing seen in multi-ensemble Markov models; the paper observes this effect but does not quantify data requirements."],"forward_implications":["Within the trained temperature range, one aTI model can produce ensembles at arbitrary target temperatures, replacing multiple dedicated low-temperature simulations.","Free-energy differences between temperatures can be estimated from very few transformed samples, which is valuable when sampling is expensive.","TI-generated equilibrium samples feed directly into gEDMD, giving temperature-dependent kinetic rates without time-correlated trajectories.","The latent TI variant shares statistical information across temperatures, so a single model can generate on-demand samples at several temperatures with improved accuracy compared with separately trained models."],"supporting_citations":[{"why":"Supplies the stochastic interpolant framework that defines the training objective for the velocity field.","marker":"[58]"},{"why":"Provides the Boltzmann generator framework and importance weights for unbiased sampling from surrogate models.","marker":"[45]"},{"why":"Targeted free energy perturbation identity that TI maps plug into for free energy estimates.","marker":"[61]"},{"why":"Generator extended dynamic mode decomposition (gEDMD) used to estimate kinetic rates.","marker":"[24]"},{"why":"Kernel-based variant of gEDMD with random Fourier features used in the kinetic analysis.","marker":"[79]"},{"why":"ChiroPaiNN architecture providing SE(3)-equivariant velocity fields for molecular systems.","marker":"[70]"},{"why":"MDQM9 dataset and surrogate-assisted MD used to generate training and reference data for the molecular systems.","marker":"[48]"},{"why":"Shows normalizing flows can skip the replica exchange ladder, motivationally underlying combining TI with parallel tempering.","marker":"[14]"}],"fun_headline_variants":["One flow map between temperatures gives thermodynamics","Thermodynamic interpolation: one model, many temperatures","Generate thermodynamics at any temperature from one flow","Flow-based sampler targets arbitrary temperatures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The kinetic rate estimates assume the molecular dynamics follows overdamped Langevin (Brownian) dynamics, so the generator used in gEDMD is an approximation for real, underdamped molecules; if that assumption fails, the predicted rates are not the true transition rates.","fun_headline_variants_meta":{"raw":{"variants":["One flow map between temperatures gives thermodynamics","Thermodynamic interpolation: one model, many temperatures","Generate thermodynamics at any temperature from one flow","Flow-based sampler targets arbitrary temperatures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001101,"raw_usage":{"total_tokens":4567,"prompt_tokens":894,"completion_tokens":3673,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":3619}},"tokens_in":510,"tokens_out":3673,"duration_ms":29460,"temperature":1.0,"reasoning_tokens":3619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:00:41.504727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run long unbiased underdamped molecular dynamics for N-methylformamide at 300 K, build a Markov state model, and compare its slowest relaxation rate with the gEDMD rate obtained from TI-generated samples at the same temperature; a substantial mismatch would localize the failure in the Brownian-generator assumption rather than the learned map.","supporting_citations":[{"cited_title":"O.; Kevrekidis, I","cited_arxiv_id":null,"evidence_quote":"Shows normalizing flows can skip the replica exchange ladder, motivationally underlying combining TI with parallel tempering."}],"review_version":1}