{"id":"f80960ce-6f9f-453c-8fea-31e4794a8901","arxiv_id":"2601.22123","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A mean-flow consistency loss lets a neural network learn large-timestep Hamiltonian flow maps from independent instantaneous phase-space samples, enabling stable MD rollouts at timesteps far beyond classical integrator limits without trajectory data.","lead":"This paper trains a neural network to predict the average phase-space motion of a Hamiltonian system over a large timestep, using only single-time samples with force labels instead of full trajectories. If it works, molecular dynamics simulations could take 10-20x larger steps and cut simulation cost while training on existing force-field datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-monotonic Δt instability in alanine dipeptide (Table 15/Fig. 22) contradicts the flow-map claim; NVE filter ablation would settle it.","rationale":"The paper contributes a clean sufficiency argument and strong empirical results: the core mathematical identity is correctly derived, and the experiments convincingly show that large-timestep rollouts can match structural observables. I was initially drawn to the reader's coverage concern (i.i.d. samples vs. trajectory data), but on close reading the more decisive issue is the non-monotonic instability in the paper's own alanine dipeptide results. A true Hamiltonian flow map—even an approximate one obtained by minimizing a residual loss—should not be unstable at Δt=10 fs and then stable again at 12 fs; the error in approximating the flow should vary smoothly. The PMF error spike (29.7 vs 0.05) is evidence that the learned field violates the consistency equation on the relevant part of phase space for those timesteps, or that the inference filter breaks down. The paper acknowledges the instability but does not resolve it; its own claim that the instability does not correlate with one-step error suggests the problem lies in the composition or the filter, not just local accuracy. This is load-bearing because the headline is stable large-timestep MD. The proposed NVE filter ablation and residual check would directly settle whether the map or the filter is at fault. The reader's verdict CONDITIONAL is appropriate; my test adds a concrete condition.","tokens_in":39842,"tokens_out":18150,"duration_ms":212257,"concrete_test":"Run NVE simulations of alanine dipeptide at Δt = 9,10,11,12,13 fs with the provided weights for 100 ns (or until collapse) under: (a) no filters, (b) drift removal only, (c) drift removal + angular-momentum filter only, (d) drift removal + coupled energy & angular-momentum filter. Track total-energy drift, angular-momentum drift, and collapse time. If collapse occurs at 10–11 fs without the coupled filter, the learned map itself is inconsistent with Hamiltonian flow at those Δt, invalidating the sufficiency claim in practice. If collapse occurs only with the coupled filter, the map may be acceptable and the filter's closed-form projection must be fixed—but then the paper's claim of stable rollouts at 12 fs is an artifact of filter/thermostat choice and must be re-evaluated. Also compute the consistency residual R = ||ūθ − (g + Δt·d/dt ūθ)|| on states collected along the 10 fs and 12 fs r","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Eqs. 6–12, Appendix C) is that minimizing the mean-flow consistency loss makes ūθ the true time-averaged Hamiltonian displacement field, so Δt·ūθ is the exact Hamiltonian flow map. A true flow map is stable for every Δt in its domain and its error grows smoothly (or at least monotonically) with Δt. Yet Table 15/Fig. 22 report a sharp spike for alanine dipeptide: PMF error 0.048 at 9 fs, 29.676 at 10 fs, 10.480 at 11 fs, then 0.053 at 12 fs. Non-monotonic stability by a factor of ~600 cannot be explained by a valid Hamiltonian flow. Either the consistency residual is large specifically in the phase-space region visited by the 10–11 fs rollouts (so the sufficiency condition is not met in practice), or the coupled energy/angular-momentum filter (Appendix B.3, especially the quadratic for λ in Eq. 51) fails at those steps. The paper states the instability does not correlate with one-step prediction error and speculates about an interaction with the simulation setup, but provides no diagnostic separating map error from filter failure. Because stable 12 fs rollouts and unstable 10–11 fs rollouts are both reported, the load-bearing claim that the learned map reconstructs the Hamiltonian flow is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Hamiltonian Flow Maps (HFMs) learned with a mean-flow consistency objective. The network ūθ(x,p,Δt) predicts the time-averaged phase-space displacement, and is trained with the consistency loss in Eq. (10) using only instantaneous labels (v,f), without trajectory pairs. The paper argues (Appendix C) that exact minimization of this loss is sufficient to recover the Hamiltonian flow map, after which rollouts are generated by x,p ← x,p + Δt·ūθ. Experiments include toy potentials, a 100-body gravitational system, MD17 small molecules, and alanine dipeptide, with stable NVT/NVE simulations at timesteps up to 9–12 fs, compared to 0.5 fs Velocity Verlet. Inference filters for drift removal, random rotation, and coupled energy/angular-momentum conservation are introduced. Code and model weights are made available.","tokens_in":40193,"tokens_out":14923,"duration_ms":156648,"significance":"If the central claim is substantiated, this is a meaningful contribution: it would allow large-timestep MD training directly on standard trajectory-free MLFF datasets, with continuous Δt conditioning and no teacher-student distillation. The derivation is clear, the experiments are broad, and the release of code and weights supports reproducibility. However, the current evidence does not fully establish that Δt·ūθ reconstructs the Hamiltonian flow map. The non-monotonic alanine dipeptide instability and the absence of reported consistency residuals leave a gap between the sufficiency proof and the empirical behavior. The empirical usefulness of the method is plausible, but the theoretical claim needs additional diagnostics before acceptance.","major_comments":[{"comment":"Table 15/Fig. 22: PMF error is 0.048 at 9 fs, 29.676 at 10 fs, 10.480 at 11 fs, and 0.053 at 12 fs. A smooth family of approximate flow maps should degrade (roughly) monotonically in Δt; this isolated spike contradicts the central claim that Δt·ūθ is the Hamiltonian flow map. The paper's 'inherent instability' remark (G.11) is not a diagnosis. Please provide: (i) one-step consistency residual and one-step rollout error vs Δt; (ii) NVE rollouts at 10/11 fs with and without the coupled E&L filter, including whether the quadratic in Eq. (51) has a real root and which root is selected; (iii) trajectory stability metrics at these timesteps.","section":"§G.11, Table 15/Fig. 22, §5.2"},{"comment":"The sufficiency proof in Appendix C assumes the loss is minimized exactly over the relevant phase-space domain. In practice the loss is evaluated on i.i.d. samples with momenta drawn from a Maxwell–Boltzmann distribution at T∼N(500 K,150 K) (Eq. 24), while alanine dipeptide is simulated at 300 K and MD17 at 500 K. The paper does not demonstrate that this training distribution covers the states visited in long rollouts, nor does it report the achieved consistency residual on those states. This is load-bearing: the trajectory-free claim requires that grafting random momenta onto static (x,f) samples is as informative as trajectory data. Please report the residual R=||ūθ−[(v,f)+Δt((v·∂x)ūθ+(f·∂p)ūθ−∂Δtūθ)]||² on the training set and on states visited during rollouts, as a function of Δt, and compare its magnitude with the one-step integration error.","section":"§4.2, Appendix B.1, Appendix C"},{"comment":"Figure 18 shows that the HFM's semigroup consistency error (u_Δt vs u_{Δt/2}∘u_{Δt/2}) grows beyond ~4 fs and is no longer comparable to Velocity Verlet at the production timesteps (9–12 fs). A true Hamiltonian flow map must satisfy the semigroup property exactly. The reported violation at the timesteps used in the main experiments indicates that the learned map deviates from the flow-map semigroup just where the method claims its advantage. Please report the semigroup error at the same Δt values as Table 15 and discuss why the violation does not degrade the equilibrium statistics, or restrict the flow-map claim to timesteps where the semigroup holds.","section":"Appendix G.8, Figure 18"}],"minor_comments":[{"comment":"'Totalling to 1 ms' should read 1 μs: 10 parallel simulations of 100 ns each sum to 1 μs, not 1 ms. Also check the unit for the source dataset description.","section":"Appendix F.5"},{"comment":"The mixture distribution q(τ) = 0.98B(1,2) + 0.02U(0,1) is written without defining the Beta distribution parameters. Specify the density or parameterization explicitly.","section":"Appendix B.1, Eq. (23)"},{"comment":"The normalization in Eq. (88) uses N as the number of integration substeps without defining it. Please define N and the indexing range.","section":"Appendix G.6, Eq. (88)"},{"comment":"Algorithm 3 includes a Thermostat step, but the NVE experiments presumably omit it. State this explicitly and clarify which filters are used in each ensemble.","section":"Algorithm 3"},{"comment":"Eq. (10) writes a plain regression loss, while Algorithm 1 uses stop-gradient on the target. Align the notation so the reader knows the target is treated as a fixed regression target.","section":"§4.2, Eq. (10) and Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the experimental scope is impressive. The main obstacle is the unexplained non-monotonic instability in the alanine dipeptide results and the lack of consistency-residual diagnostics. If the requested experiments show that the instability is a filter/thermostat artifact and the residual is small on rollout states, I would be willing to support acceptance after a minor revision. If not, the flow-map claim should be substantially softened, even though the method may still be useful as a heuristic large-timestep sampler."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The first thing to know: this paper is worth a serious referee. It adapts the mean-flow consistency condition from generative modeling to deterministic Hamiltonian dynamics, so you can train a large-timestep integrator from single-time snapshots (positions + forces) without generating trajectories. That is a real step beyond Bigi, Thiemann, and the other trajectory-based methods, and the experiments back it up on toy systems, N-body, MD17, and alanine dipeptide. The model also works as a conventional force field, and it survives on 256 training samples for paracetamol, which is impressive.\n\nThe math in App C is sound under exact loss minimization. The target in Eq. (10) is self-referential — it uses the network's own Jacobian-vector products with stop-gradient — but the Δt=0 anchor is grounded in external force labels, and the rollouts are checked against analytical potentials and ab initio reference distributions. So the circularity is real but contained.\n\nNow the soft spot. The paper reports a non-monotonic instability for alanine dipeptide: PMF error is 0.048 at 9 fs, 29.7 at 10 fs, 10.5 at 11 fs, then 0.053 at 12 fs. That is not what a valid Hamiltonian flow map does. Either the consistency residual is large specifically in the phase-space region visited by those rollouts, or the coupled energy/angular-momentum filter (the quadratic for λ in Eq. 51) breaks down at those steps. The paper acknowledges the instability but does not separate map error from filter failure. That is the load-bearing question, and the current text hand-waves it with 'interaction between the learned map and the simulation setup.'\n\nRelated: the stability claims lean on post-hoc filters (random rotation, drift removal, coupled conservation projection). The ablation shows the filters do not rescue Velocity Verlet, which is good evidence the learned map is doing real work. But we still need an HFM rollout with the energy/angular-momentum filter disabled to see how far the map alone goes. The NVE validation is also thin: 500 ps for one molecule.\n\nThe other weakness is the trajectory-free premise itself: momenta are sampled from a broad Maxwell-Boltzmann at inference-agnostic temperatures, and the claim is that this covers the joint distribution visited during long rollouts. That is plausible but not demonstrated.\n\nBottom line: send it to review. The idea is genuinely new, the experiments are broad, and the limitations are disclosed rather than hidden. Ask the authors to (i) diagnose the 10–11 fs spike with a filter ablation and per-Δt residual analysis, and (ii) show NVE stability and filter-free behavior for at least one more system. If those hold up, this is a solid contribution to the MLFF/large-timestep literature.","headline":"Trajectory-free consistency training for large-timestep MD is a real new idea with broad experiments, but the unexplained 10-11 fs instability in alanine dipeptide prevents accepting the flow-map claim as-is.","tokens_in":40741,"tokens_out":3425,"would_cite":true,"duration_ms":32803,"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":"A neural network can learn stable molecular-dynamics integrators at 9–12 fs timesteps from instantaneous forces alone, with no trajectory data, by enforcing a mean-flow consistency identity on the time-averaged dynamics.","keywords":["Hamiltonian flow maps","mean flow consistency","large-timestep molecular dynamics","trajectory-free training","machine-learned force fields","consistency training","molecular dynamics acceleration","phase-space integration"],"falsifier":"Measure single-step flow-map error (RMSD of predicted (x,p) vs. a fine-grained reference integration) on states drawn from a long rollout of the model itself — states trajectory-free training never saw — and compare with the error on decorrelated training-like states. The paper reports low normalized one-step error on training-like states (Appendix G.6) yet an instability band near Δt ≈ 10 fs for alanine dipeptide (Appendix G.11); that gap is exactly what this test resolves. If rollout-state error stays low at 9 fs, the sufficiency claim survives; if it balloons while training-state error is t","tokens_in":39724,"feed_emoji":"⚛️","tokens_out":13751,"duration_ms":139208,"temperature":0.7,"pith_summary":"The paper sets out to break the timestep bottleneck of molecular dynamics: classical integrators such as Velocity Verlet need steps of about 0.5 femtoseconds, which makes long simulations extremely expensive. The authors claim that a network can instead learn the Hamiltonian flow map — the rule that advances positions and momenta by one large interval — without ever seeing a trajectory. The key move is to learn the time-averaged displacement field over the interval and to enforce a 'mean flow consistency' identity that this field must satisfy, an identity built only from the instantaneous velocity and force at the starting point. Because the constraint is local in time, the model trains on exactly what standard machine-learned force-field datasets already contain: decorrelated geometries with force labels, plus randomly sampled momenta. If the claim holds, any MLFF can double as a large-step integrator at comparable training cost, with one model predicting both instantaneous forces and multi-femtosecond jumps.","feed_headline":"18x larger molecular-dynamics timesteps from forces alone","feed_subtitle":"No trajectory data needed: a mean-flow consistency identity turns force-field labels into a large-step integrator.","key_machinery":"The central object is the startpoint-conditioned mean displacement field ū(x,p,Δt), the time-averaged velocity and force accumulated along a Hamiltonian trajectory over an interval Δt. The load-bearing identity is the forward consistency condition obtained by differentiating the integral that defines ū with respect to its start time: ū = (v,f) + Δt[(v·∂x)ū + (f·∂p)ū − ∂Δt ū]. This converts a nonlocal time integral into a local differential constraint that can be imposed as a regression loss on individual phase-space samples, anchored at Δt=0 by the instantaneous force. The flow map is recovered as (x,p) + Δt·ū, and a closed-form coupled conservation filter — solving a quadratic equation in o","core_discovery":"A neural network ūθ trained to minimize the mean-flow consistency loss L = E_Δt‖ūθ − [(v,f) + Δt((v·∂x)ūθ + (f·∂p)ūθ − ∂Δt ūθ)]‖², evaluated at single phase-space points with only instantaneous labels (v,f), learns the true time-averaged Hamiltonian displacement field. Appendix C shows that any field satisfying this forward consistency identity obeys the same equation as the genuine interval average, and the two coincide at Δt = 0. The flow map follows by multiplication, (x*,p*) = (x,p) + Δt·ū; one model thus predicts forces at Δt=0 and advances simulations by 9–12 fs per step — versus 0.5 fs for Velocity Verlet — while recovering equilibrium statistics, vibrational spectra, and the referenc","pith_inferences":["The derivation never uses trajectory pairs, so the same local consistency identity should transfer to systems where instantaneous derivatives are cheap but trajectories are not — non-separable Hamiltonians, learned coarse-grained equations, or stochastic dynamics with known drift — with the flow-map reconstruction unchanged.","Because the map is only approximately symplectic, the largest stable Δt should track the system's Lyapunov time; a testable prediction is that the maximal usable step shrinks as the local divergence rate grows, consistent with the paper's own report that optimization degrades in strongly chaotic regimes and that alanine dipeptide has an instability band around Δt ≈ 10 fs (Appendix G.11).","The biggest gap between the proof and the practice is the grafting of Maxwell–Boltzmann momenta onto decorrelated geometries: comparing rollout stability of a model trained that way against one trained on trajectory-consistent (x,p) pairs would directly measure how much the procedure loses by discarding temporal correlations."],"forward_implications":["Stable NVT and NVE rollouts at Δt = 9–12 fs for small molecules and peptides, roughly 18–24 times fewer integration steps than a 0.5 fs Velocity Verlet baseline.","Trajectory-free training: standard MLFF datasets of decorrelated geometries with force labels suffice, with momenta sampled from a Maxwell–Boltzmann distribution.","One model serves two roles — conventional force field at Δt = 0 and a continuous-time large-step integrator for any Δt in [0, Δt_max] — so no per-timestep retraining is needed.","Compute stays comparable to standard MLFF training (about 30% overhead), and simulated throughput reaches hundreds of ns per GPU-day at the largest steps.","Learned maps reproduce both static observables (interatomic distance distributions, free-energy surfaces, metastable basin transitions) and dynamical ones (vibrational power spectra)."],"fun_headline_variants":["18x MD timestep from force labels alone","Large MD steps without trajectory data","Mean-flow consistency unlocks big timesteps","Force-only training: 18x faster MD simulation","Hamiltonian flow maps: 12 fs steps, no trajectories"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof that the consistency loss suffices assumes it is minimized to zero (Appendix C), and the trajectory-free training premise assumes that decorrelated geometries with independently sampled Maxwell–Boltzmann momenta (Section 4.2) cover the phase-space region a long rollout actually visits — so a field made consistent at those training points stays correct along the simulated trajectory.","fun_headline_variants_meta":{"raw":{"variants":["18x MD timestep from force labels alone","Large MD steps without trajectory data","Mean-flow consistency unlocks big timesteps","Force-only training: 18x faster MD simulation","Hamiltonian flow maps: 12 fs steps, no trajectories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1363,"prompt_tokens":709,"completion_tokens":654,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":582}},"tokens_in":453,"tokens_out":654,"duration_ms":7089,"temperature":1.0,"reasoning_tokens":582,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:44:10.477078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure single-step flow-map error (RMSD of predicted (x,p) vs. a fine-grained reference integration) on states drawn from a long rollout of the model itself — states trajectory-free training never saw — and compare with the error on decorrelated training-like states. The paper reports low normalized one-step error on training-like states (Appendix G.6) yet an instability band near Δt ≈ 10 fs for alanine dipeptide (Appendix G.11); that gap is exactly what this test resolves. If rollout-state error stays low at 9 fs, the sufficiency claim survives; if it balloons while training-state error is t","supporting_citations":[],"review_version":1}