{"id":"488da8a2-f76f-4ef2-876f-ca2a96ca7039","arxiv_id":"2505.02450","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MDPNet uses a multiscale diffusion autoencoder and a graph neural ODE to predict spatiotemporal dynamics, reporting 53.23% average error reduction on four PDE benchmarks.","lead":"This paper introduces a neural network that predicts how complex physical systems evolve by learning separate coarse-to-fine descriptions of the system and how those descriptions influence each other. It reports roughly 53% lower prediction error than baseline models on four benchmark systems, including fluid flows and chemical reactions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 53.23% headline reduction is dominated by the Brusselator result, and the paper's evaluation wording leaves open whether Bruss metrics are computed only on a post-transient limit-cycle segment; this should be checked before accepting the headline.","rationale":"The reader's conditional verdict is appropriate: the architecture is plausible, the code is released, and the experiments span four systems, but the headline number was not independently reproduced. My stress-test focuses on the most load-bearing empirical quantity, the 53.23% average reduction. The table arithmetic checks out, but the average is dominated by the Bruss result, and the paper's own appendix introduces ambiguity about whether Bruss is evaluated in the periodic limit-cycle regime. If the limit-cycle selection is only for snapshots, the concern would weaken; if it also applies to metric computation, the central claim is materially overstated. This is directly settleable by running the released code with full-trajectory metrics. I do not see an internal inconsistency in the model equations themselves, though the reader's latent-Markov concern remains a secondary theoretical risk: if the encoded z is not a sufficient statistic, the GNODE could drift, but that risk is best assessed through the benchmark numbers and rollouts. I therefore keep the CONDITIONAL verdict and recommend the protocol check as the condition for acceptance.","tokens_in":19822,"tokens_out":9621,"duration_ms":120357,"concrete_test":"Run the released code's Brusselator evaluation so that NMSE and SSIM are computed over the full 100-step trajectory from the initial condition, including the transient, and separately over the post-transient limit-cycle segment; then recompute the four-system average using a fixed baseline (e.g., FNO or ConvLSTM) rather than the per-system best baseline. If the full-trajectory Bruss NMSE rises above roughly 0.5e-2 or the alternative average falls materially below 53.23%, the headline is an artifact of the evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the 53.23% average NMSE reduction in Table 1. That number is the mean of four per-system reductions, each computed against a different best baseline: LO 12.40% vs G-LED, Bruss 95.03% vs ConvLSTM, GS 63.30% vs L-DeepONet, NS 42.18% vs ConvLSTM. The Bruss case alone contributes 95.03%, and excluding it the average over the other three systems is 39.29%. The Bruss result is therefore load-bearing for the headline. The paper's own dataset description says the Brusselator converges to a limit cycle after an initial transient, and Appendix B.1 states that for the Bruss system 'a segment of a limit cycle after long-term evolution is selected' when reporting prediction snapshots. If the same selection was used for the metrics in Table 1, the test reduces to forecasting a known periodic orbit from within the periodic regime, which is substantially easier than the stated problem of predicting from an arbitrary initial condition (Sec. 2.1). The discrepancy between the 0.044e-2 Bruss NMSE and the next-best 0.886e-2 is large enough that this protocol choice, rather than the architecture, could explain the reported 95% improvement. Because the text is ambiguous about whether the limit-cycle selection applies only to visualizations or also to metric computation, the headline is not yet robustly supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MDPNet, a multiscale diffusion autoencoder with an attention-based graph neural ODE for predicting the spatiotemporal dynamics of complex systems. The method encodes multiscale residual features, uses a coarsening-guided diffusion decoder to reconstruct observations, and models cross-scale co-evolution in a low-dimensional latent space. The authors evaluate on four PDE systems (Lambda-Omega, Brusselator, Gray-Scott, Navier-Stokes) and report an average NMSE reduction of 53.23% against the best baseline per system, along with robustness, generalization, and ablation experiments.","tokens_in":20163,"tokens_out":6663,"duration_ms":71311,"significance":"If the results hold, the paper offers a novel and well-motivated architecture that explicitly exploits multiscale structure in a diffusion autoencoder, with a released codebase and a fairly comprehensive empirical study across four PDE benchmarks, sensitivity analyses, ablations, robustness to noise, and out-of-distribution generalization. The key claim, however, hinges on the Brusselator result, which contributes 95.03% of the average improvement, and the evaluation protocol for that system is ambiguous. Because the headline number depends on a single system and an unclear test segment, the current evidence does not yet robustly support the paper's central quantitative claim.","major_comments":[{"comment":"The evaluation protocol for the Brusselator is ambiguous and load-bearing. The dataset description states that the Brusselator converges to a limit cycle after a transient, and Appendix B.1 says that \"for the Bruss system, a segment of a limit cycle after long-term evolution is selected\" when reporting prediction snapshots. The main text does not state whether the NMSE/SSIM metrics in Table 1 are computed over full trajectories from arbitrary initial conditions (as implied by Sec. 2.1 and Sec. 4.3) or over a post-transient limit-cycle segment. If the metrics are computed only on the periodic segment, then the 95.03% reduction over ConvLSTM is not evidence of a general prediction capability from arbitrary initial conditions, and the 53.23% average should be recomputed or explicitly scoped. The authors must specify exactly which segment was used for Table 1 and ensure all baselines are evaluated identically.","section":"Sec. 4.1, Appendix B.1, Table 1"},{"comment":"The definition of the coarsening-guided diffusion schedule is incomplete and the forward process is not a standard diffusion process. The schedule is given as {N_k}_{k=1}^{K-1}, but the stage boundaries N_0 and N_K are undefined, and Eq. (5) sets alpha=1 for n<N_k, making the forward variable a coarse-grained deterministic transform rather than a noisy sample. The reverse process \"starts from Gaussian noise,\" but the correspondence between the initial reverse sample and the forward endpoint is not specified. This makes the training objective in Eq. (10) ambiguous regarding the range of n. The authors should provide a complete, self-consistent specification of the forward and reverse processes, including the role of Q and the exact values of N_k.","section":"Sec. 3.1.2, Eqs. (5)-(7)"},{"comment":"The headline claim of an \"average prediction error reduction of 53.23%\" is the mean of per-system improvements, each computed against a different best baseline (G-LED for LO, ConvLSTM for Bruss, L-DeepONet for GS, ConvLSTM for NS). This is not a single consistent comparison. The Bruss result alone contributes 95.03 percentage points of that average; excluding it, the mean over the other three systems is 39.29%. The abstract and conclusions should either report the per-baseline reductions or state clearly that the average is over system-specific best baselines, and the paper should discuss the sensitivity of the headline number to the Bruss protocol.","section":"Abstract, Table 1"}],"minor_comments":[{"comment":"The text says \"remains theoretically lossless when k=1 (i.e., x_tau = x^1_tau)\", but by Eq. (4) x^1_tau is the sum of residuals from scales 2..K and the residual r^1_tau would still be nonzero. The correct statement is x_tau = x^1_tau + r^1_tau. Please fix the typo.","section":"Sec. 3.1.1, Eq. (4)"},{"comment":"The table columns labeled \"Diffuse Steps\" and \"Scale Num\" are not explicitly connected to the control experiment described in the text. The left panel appears to be the single-scale control with varying total diffusion steps, and the right panel the multiscale model with fixed 200 steps per scale. Please label these panels in the table or in the caption.","section":"Sec. 4.4.2, Table 2"},{"comment":"In the table for the cylinder flow system, the entry in the alpha column appears as \"1.0μ\" and the viscosity range is rendered as \"1.06×10 4\"; these are formatting errors. Also, the text says \"We generate 50 training trajectories and 20 testing trajectories\" but then refers to \"10 training trajectories\" and \"10 OOD\" trajectories; the numbers should be made consistent.","section":"Appendix A, Table 1"},{"comment":"The model assumes that the latent state z^k_tau is a sufficient statistic for future dynamics under a first-order Markovian ODE. For systems where the coarse-graining operator Q discards information needed to determine the future, or where memory beyond the current latent vector is required, the latent dynamics will drift. This limitation is not discussed; a sentence acknowledging it would improve the paper.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong empirical systems paper, and the architecture is interesting. The main risk is the Brusselator evaluation protocol: if the metrics are computed on a post-transient limit-cycle segment, the headline 53.23% reduction is not a fair comparison to the stated problem of predicting from an arbitrary initial condition. I would recommend the editor ask the authors to clarify this before any final decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper is a solid, competent empirical method paper, not a breakthrough. The architecture is a real new combination: a multiscale residual encoder, a coarsening-guided diffusion decoder, and an attention-based graph neural ODE for cross-scale latent dynamics. Each piece has prior art, but the wiring together is new, and the ablation study earns the design. The paper also ships code, includes robustness and generalization experiments, and acknowledges its grid-based limitation. That is more than most submissions do.\n\nThe main thing to check is the headline. The 53.23% average reduction is the mean of four per-system improvements, each computed against a different best baseline. That is promotional but not wrong. Excluding the Brusselator, the average over the other three systems is about 39.3%, which is still respectable. The Bruss number itself is arithmetically consistent with the table: (0.886-0.044)/0.886 = 95.0%, so the 'inconsistent with NMSE' reading from the stress-test is off. The real soft spot is the evaluation protocol. Appendix B.1 says that for Bruss, 'a segment of a limit cycle after long-term evolution is selected' for prediction snapshots. If the same selection is used when computing Table 1 metrics, the test becomes 'forecast a known periodic orbit from inside the periodic regime,' which is much easier than the stated problem of predicting from an arbitrary initial condition. The paper doesn't say which one it is. That ambiguity is load-bearing for the headline and needs to be resolved before I'd trust the 95% claim.\n\nLess serious: the latent dynamics are assumed first-order Markovian in Eq. 8, and if the coarse-graining operator discards information needed for forecasting, the latent ODE will drift. That's a theoretical caveat, not a demonstrated failure. The residual decomposition in Eq. 3-4 is fine; the main mathematical risk is the Markov assumption. Hyperparameters like scale number and diffusion-step allocation are tuned on the same benchmarks; that's mild, not circular. The citation pattern looks fair, and the code link is a real plus.\n\nBottom line: this paper deserves a serious referee. I'd send it out, with the request that the authors state explicitly whether Table 1 metrics are computed on full trajectories or only the limit-cycle segment of Bruss, and ideally report a common-baseline aggregate. If that comes back clean, accept.","headline":"Solid new architecture and broad experiments, but the 53.23% headline leans heavily on the Brusselator result and the test protocol needs clarification before I'd trust it.","tokens_in":20659,"tokens_out":6405,"would_cite":true,"duration_ms":69950,"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":"MDPNet predicts complex-system dynamics by separating states into coarse-to-fine scales, guiding a diffusion decoder with scale-specific codes, and coupling scales through a graph neural ODE, cutting average prediction error by 53.23…","keywords":["complex systems","multiscale modeling","diffusion models","latent dynamics","graph neural ODE","spatiotemporal prediction","PDE systems","coarse-graining"],"falsifier":"A concrete test: use a system whose long-term fate is set by small-scale initial perturbations, so two states that look identical after downsampling evolve differently. If MDPNet's latent vectors do not distinguish them, its forecast error will jump relative to a full-resolution model, showing that the Markovian latent state alone is insufficient.","tokens_in":19624,"feed_emoji":"🌊","tokens_out":8933,"duration_ms":100296,"temperature":0.7,"pith_summary":"Complex systems like fluids and reacting chemicals evolve on many scales at once, and this paper argues that a predictor ignoring that structure leaves accuracy on the table. It proposes MDPNet, a forecasting model whose autoencoder deliberately separates a state into coarse-to-fine residual scales, whose diffusion decoder is conditioned on those scale-specific codes, and whose predictor is a graph neural ordinary differential equation coupling the scales' latent dynamics. On four standard partial differential equation systems, the paper reports an average 53.23% reduction in normalized mean-squared error against ten baselines, plus improved robustness to noise and data scarcity and better extrapolation to unseen parameter regimes. A sympathetic reader would take the contribution to be the demonstration that the latent space of a complex system should be multiscale by construction rather than a single compressed vector.","feed_headline":"Splitting states into scales cuts forecast error 53%","feed_subtitle":"A multiscale diffusion autoencoder with graph-ODE coupling beats ten baselines on reaction-diffusion and fluid systems.","key_machinery":"The central machinery is a multiscale diffusion autoencoder paired with a graph neural ODE over scales. The encoder decomposes the state $x_t$ into residuals $r^k_t = Q(x_t - \\sum_{i>k} r^i_t, k)$, where $Q(\\cdot,k)$ is a downsampling operator followed by interpolation back to the original resolution; the decoder allocates diffusion stages to each scale and conditions the noise network on $z^k_t=\\phi_\\theta(r^k_t,\\mathrm{embedding}(k))$, so coarse-to-fine scale structure is what guides denoising. The predictor evolves $z^k_t$ through self-dynamics $f_k$ plus a graph attention interaction term $g$ that reads the full latent vector $\\mathbf{Z}_t$ and the adjacency $A$, solving an initial value problem with an ODE solver. What this machinery does is keep fine-scale information explicitly represented and let the forecast update each scale using the others' states, rather than compressing everything into one entangled code.","core_discovery":"The paper's central discovery is that decoupling observed states into several scales and guiding a diffusion decoder with those scale-specific latent vectors yields a latent space that supports markedly better forecasts than flat latent spaces. The residual encoder computes coarse-grained states by repeatedly extracting and subtracting downsampled versions of the state, so that finer scales carry what coarser scales discarded; the diffusion decoder's noise schedule is staged per scale, applying coarse-graining instead of pure noise in early steps, and each stage is conditioned on the corresponding latent vector. The predictor then treats scales as nodes of a fully connected graph and evolves them through an attention-based graph neural ODE, $\\frac{dz^k_t}{dt}=f_k(z^k_t)+g(z^k_t,A,\\mathbf{Z}_t)$. In experiments, this combination outperforms all baselines on the Lambda-$\\Omega$, Brusselator, Gray-Scott, and Navier-Stokes systems, with the largest gains on Brusselator and Gray-Scott, and the diffusion decoder tolerates latent perturbations better than a vanilla residual decoder, which the paper ties to smaller long-horizon error accumulation.","pith_inferences":["Editorial inference: because the coarse-graining operator $Q$ defines the decomposition, the same autoencoder could be lifted to graphs or point clouds by swapping $Q$, which the paper's stated limitation (grid-only data) invites as the next test.","Editorial inference: the $K=3$ sweet spot in the scale-number experiments suggests the optimal number of scales tracks the number of dynamically active structure sizes, so a data-driven rule could choose $K$ from the spectral decay rate of residuals rather than by grid search.","Editorial inference: the learned graph attention weights over scales can be read as a per-system measure of cross-scale coupling; correlating them with known energy-cascade directions in Navier-Stokes would test whether the model rediscovers physical cascade structure."],"forward_implications":["Removing the cross-scale interaction term from the graph neural ODE substantially raises NMSE on the Lambda-Omega and Brusselator systems, so the co-evolution model carries a large share of the accuracy gain.","Because the decoder conditions on latent codes rather than reconstructing residuals and summing them, latent perturbations cause smaller reconstruction drift than a vanilla decoder, implying that long-horizon forecasts are less vulnerable to error accumulation.","On the Brusselator system, MDPNet trained on only 60% of the data still beats the best baseline trained on the full set, so the multiscale inductive bias improves data efficiency.","On cylinder flow with Reynolds numbers between 500 and 1000, outside the training range, MDPNet still reproduces vortex count, shape, and position, so the latent representation generalizes beyond the training parameter regime.","Forecasting runs at $K \\times d$ latent dimensions instead of $C\\times H\\times W$ (over a 40-fold reduction on the cylinder flow), with denoising steps independent of forecast horizon, so the method is positioned as a cheaper surrogate for numerical simulation."],"supporting_citations":[{"why":"Supplies the denoising diffusion objective and reverse-process sampler the decoder is trained with.","marker":"[23]"},{"why":"Supplies the latent-space forecasting baseline (AE-LSTM) and the coarse-to-fine influence assumption the encoder builds on.","marker":"[68]"},{"why":"Supplies the diffusion-reconstruction latent baseline (G-LED) that MDPNet must beat and the subsampling idea it extends.","marker":"[16]"},{"why":"Supplies the graph neural ODE formulation the cross-scale predictor is built from.","marker":"[79]"},{"why":"Supplies the graph attention layer used in the cross-scale interaction term.","marker":"[65]"},{"why":"Grounds coarse-graining of PDEs as a route to low-dimensional effective dynamics, the physical idea behind the residual encoder.","marker":"[1]"},{"why":"Supplies the multiscale diffusion schedule idea that motivates stage-wise coarse-graining in the decoder.","marker":"[14]"},{"why":"Supplies the neural ODE solver used for latent evolution and the Latent ODE baseline.","marker":"[10]"}],"fun_headline_variants":["Multiscale diffusion autoencoder predicts complex dynamics with 53% less error","Graph ODE plus multiscale diffusion slashes prediction error by half","Scale-splitting autoencoder and graph ODE cut forecast error 53%","Multiscale diffusion meets graph ODE: 53% lower prediction error","Multiscale diffusion autoencoder achieves 53% better complex dynamics forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the current multiscale latent vector is enough to forecast the next state; if the true dynamics depend on fine details that downsampling throws away or on history beyond the present instant, the predicted states will drift.","fun_headline_variants_meta":{"raw":{"variants":["Multiscale diffusion autoencoder predicts complex dynamics with 53% less error","Graph ODE plus multiscale diffusion slashes prediction error by half","Scale-splitting autoencoder and graph ODE cut forecast error 53%","Multiscale diffusion meets graph ODE: 53% lower prediction error","Multiscale diffusion autoencoder achieves 53% better complex dynamics forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":2987,"prompt_tokens":927,"completion_tokens":2060,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1960}},"tokens_in":543,"tokens_out":2060,"duration_ms":14023,"temperature":1.0,"reasoning_tokens":1960,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:51:18.698163+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: use a system whose long-term fate is set by small-scale initial perturbations, so two states that look identical after downsampling evolve differently. If MDPNet's latent vectors do not distinguish them, its forecast error will jump relative to a full-resolution model, showing that the Markovian latent state alone is insufficient.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the latent-space forecasting baseline (AE-LSTM) and the coarse-to-fine influence assumption the encoder builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-reconstruction latent baseline (G-LED) that MDPNet must beat and the subsampling idea it extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the graph neural ODE formulation the cross-scale predictor is built from."}],"review_version":1}