{"id":"22109d05-dbf6-4902-a8b2-4a95c45d61e7","arxiv_id":"2506.11328","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"ASNO combines a transformer with a nonlocal attention operator under a BDF-inspired decomposition and claims state-of-the-art generalization, but the evaluation is undermined by an uneven comparison where ASNO receives future forcing inputs.","lead":"This paper proposes ASNO, which splits predictions of physical systems into a temporal forecast and a spatial correction, following a classical numerical scheme. On benchmark tests it reports better accuracy and generalization than standard models, but the comparison appears to give ASNO access to future forcing values, making the headline result unreliable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ASNO is fed future forcing F_{m+1} while the baselines receive only F_m, so the reported accuracy and OOD gains may be an information advantage rather than an architectural one.","rationale":"The paper's strongest claim is that ASNO outperforms existing models in accuracy, long-term stability, and zero-shot generalization. The most load-bearing condition for that claim is a fair comparison. The manuscript explicitly gives ASNO the future forcing field F_{m+1} (Section 3.4, Eq. 17; Figure 1) while stating that all baseline models use the current forcing field F_m (Section 4.1; Section 4.2). This is an input-information asymmetry, not a difference in inductive bias or parameterization. Since ASNO's inputs are a superset of the baselines' inputs, the lower losses in Tables 1 and 2 cannot be attributed to the architecture. The concern is concrete and testable: switching ASNO to F_m, or giving F_{m+1} to baselines, should be a one-line data-format change. No part of the paper reports such a controlled comparison. The BDF decomposition could still be a valid modeling idea, and the physics-interpretability claims are interesting, but the empirical foundation of the central claim is not established as stated. For that reason the reader's REJECT verdict is appropriate, and the recommended verdict remains REJECT.","tokens_in":14692,"tokens_out":4229,"duration_ms":46339,"concrete_test":"Retrain ASNO from Eq. (17) with F_m in place of F_{m+1} on the Darcy setup (Table 1) and the Lorenz setup (Table 2), keeping all hyperparameters, data splits, and training procedures fixed. If ASNO's best and OOD losses rise to the level of Transolver or GNOT, the reported superiority is an artifact of future-force information. A complementary check is to feed F_{m+1} to all baselines under the same protocol; if their losses drop accordingly, the asymmetry is confirmed. This is a direct ablation requiring only the input tensor to be shifted by one time index.","verdict_should_be":"REJECT","load_bearing_attack":"The central performance claim is that ASNO achieves the lowest test and OOD losses (Section 4.1, Tables 1–3). For that claim to be supported, ASNO and the baselines must be compared under the same input information. The paper violates this condition: Eq. (17) and Figure 1 define ASNO as ASNO(X_{m-4},...,X_m, F_{m+1}), while Section 4.1 states that all baselines receive the past five states and the current forcing field F_m, and Section 4.2 repeats that each model receives the five previous states plus the current forcing F_m. Thus ASNO has direct access to the external loading at the time step it must predict, and the baselines do not.\n\nThis is not a minor implementation detail. In the Darcy OOD-f experiment the source is time-varying (g(t,x)=g~(x) sin(t)), so F_{m+1} carries information about the next forcing state that F_m does not; in the Lorenz experiments the forcing terms g_i(t) are time-dependent in the same way. Consequently, the reported margin over Transolver, GNOT, and other baselines could be explained entirely by this input asymmetry. The paper provides no ablation in which ASNO is restricted to F_m, nor an experiment in which baselines are allowed F_{m+1}. The interpretability and physics-discovery results are also tied to the same trained model, so they inherit the same concern. The BDF-style decomposition in Eq. (17) is an inductive bias that might be reasonable, but the empirical comparison does not isolate its contribution because the information available to the competing models differs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ASNO, a spatio-temporal neural operator that combines a Transformer encoder for temporal extrapolation with an attention-based nonlocal operator for spatial correction, inspired by the implicit-explicit BDF scheme. The authors claim that ASNO generalizes zero-shot to unseen physical parameters, outperforms existing neural operators and transformers on Darcy flow, Lorenz, Navier–Stokes, and additive manufacturing benchmarks, and yields interpretable components that match a BDF5 extrapolation and a Darcy-derived kernel. The central architectural idea is interesting, but the empirical evaluation has a serious input-information asymmetry: ASNO receives the future forcing F_{m+1}, while all baselines receive only the current forcing F_m. Additional issues include overstated OOD claims, inconsistent dataset arithmetic in the Lorenz experiment, and the absence of error bars. The current evidence does not support the paper's main claims.","tokens_in":15085,"tokens_out":5897,"duration_ms":62028,"significance":"If the claims were supported, ASNO would be a meaningful contribution: a separable, BDF-inspired architecture that isolates temporal and spatial effects, with potential for interpretability and zero-shot generalization in scientific ML. The paper also addresses a practically relevant problem (forecasting with known future loadings, e.g., in additive manufacturing). The architecture itself is plausible and the experimental breadth is commendable. However, the significance is currently undercut by the comparison protocol: because ASNO is given F_{m+1} while baselines are not, the reported accuracy and OOD gains could stem entirely from this information advantage. The interpretability results are also post-hoc and not quantitatively validated. The paper would be significant only after a fair comparison and a more careful treatment of OOD claims.","major_comments":[{"comment":"The central comparison is unfair. Eq. (17) defines ASNO(X_m,...,X_{m-n+1}, F_{m+1}), and Figure 1 shows F_{m+1} as an input, but Section 4.1 states that all baselines receive the past five states and the current forcing field F_m; Section 4.2 repeats that each model receives the five previous states plus F_m. Thus ASNO has direct access to the forcing at the time step it must predict, whereas the baselines do not. In the Darcy OOD-f experiment the forcing is time-varying (g(t,x)=g~(x) sin(t)), so F_{m+1} carries information that F_m does not. The reported margins over Transolver, GNOT, and other baselines in Tables 1–3 could therefore be explained by input asymmetry rather than architectural merit. The paper provides no ablation with ASNO restricted to F_m and no experiment giving F_{m+1} to baselines. This must be fixed before any performance claim can be evaluated.","section":"§4.1, Eq. (17); §4.2"},{"comment":"The paper repeatedly claims zero-shot generalization to unseen initial conditions (e.g., the abstract and contribution list), but no experiment with unseen initial conditions is reported. The Darcy OOD tests vary the forcing distribution and the permeability field, the Lorenz experiment uses the same initial condition (0,1,0) throughout, and the Navier–Stokes section reports only test loss without any OOD shift. Either an explicit out-of-distribution initial-condition experiment must be added, or the claim must be revised to refer only to the tested generalization modes.","section":"§1 and §4.1–4.3"},{"comment":"The Lorenz dataset arithmetic is inconsistent. The text says there are 2000 profiles from 100 coefficient combinations times 20 loading functions, with 80% for training and 20% for testing. That implies 1600 training profiles, yet the calculation given is '80×995×20 = 1,592,000 training samples', which uses 80 profiles, not 1600. If the intended split is 1600 profiles, the number of training samples should be 1600×995×20 = 31,840,000 before any subsampling; if the intended split is indeed 80/20 profiles, the statement that the dataset has 2000 profiles is wrong. This makes the reported sample counts and the resulting test-loss numbers unreliable.","section":"§4.2, Lorenz dataset"},{"comment":"All performance tables report only the 'Best Test Loss' over what appears to be a single run or a selected checkpoint. No standard deviations, no seed counts, and no training/validation protocol are given. Without repeated seeds or error bars, the claim that ASNO 'achieves the lowest' loss and 'outperforms' baselines is not statistically supported, especially where margins are small (e.g., Lorenz: 0.000794 vs 0.000835 for Transolver). The paper should report mean and standard deviation over multiple random seeds and a defined model-selection rule.","section":"Tables 1–3"},{"comment":"The interpretability results are post-hoc and not quantitatively validated. The claim that the Transformer latent 'closely matches' BDF5 and that the NAO kernel 'agrees with' the Darcy kernel is based on visual comparison only; no error metric, no confidence interval, and no test on out-of-distribution systems is provided. Moreover, the NAO kernel is compared with a kernel derived from a specific Darcy discretization using the same equation family on which the model was trained, so the alignment may reflect memorization of the training distribution rather than discovery of a general law. A quantitative comparison and a held-out test are needed before the physics-discovery claim can be supported.","section":"§4.1, Eqs. (19)–(22), Figures 2–3"}],"minor_comments":[{"comment":"The Lorenz section says 'We did not include architectures such as U-Net, FNO, and GNOT in our Lorenz comparison', but Table 2 includes GNOT. This contradiction should be resolved.","section":"§2.2 and Table 2"},{"comment":"The list of baselines in the text mentions 'Transformer Encoder + NAO', but the table lists 'Linear Enc. + NAO' instead. Clarify which baseline was used and whether the naming is consistent.","section":"§4.1, Table 1"},{"comment":"Equation (19) writes the BDF approximation as \\tilde X_m, while the text and Figure 2 compare the latent H_{m+1} with \\tilde X_{m+1}; the index shift should be made explicit, and the formula should be written for the same time index as the comparison.","section":"Eq. (19) and Figure 2"},{"comment":"There are several grammatical and typographical issues, e.g., 'ASNO outperforms over existing models' in the abstract and 'outpertforms' in the conclusion. The paper should be proofread.","section":"Abstract and §5"},{"comment":"The additive manufacturing experiment reports a single MAPE value of 2.50% with no comparison to baselines, no error bars, and no description of how the MAPE is computed across the 290,400 test samples. Include a table or comparison to make the result interpretable.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The input-information asymmetry alone is sufficient to invalidate the current performance comparison. The issue is fixable by re-running experiments with matched inputs (either giving F_{m+1} to baselines or restricting ASNO to F_m), but the revision must be substantial. If the authors cannot provide a fair comparison, the paper should be rejected. The OOD and interpretability claims also need to be either supported by appropriate experiments or substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — this one has a good architectural instinct and an evaluation that doesn't support its claims. The key problem: ASNO is given the future forcing F_{m+1} (Eq. 17, Fig 1), while every baseline is fed only the current forcing F_m (Sections 4.1, 4.2). For the OOD-f Darcy case the forcing is time-varying, g(t,x)=g~(x) sin(t), so F_{m+1} is genuine next-loading information the baselines never see. That alone could explain the reported margin. There is no ablation on either side of the information boundary. This is a load-bearing flaw, not a quibble.\n\nWhat is genuinely new: the BDF/IMEX-inspired decomposition into a transformer that does multistep temporal extrapolation and an attention-based nonlocal operator that applies the spatial/forcing correction is a fresh way to structure a surrogate. Pairing that explicit inductive bias with these two modules is new relative to the cited literature, and the paper's range of benchmarks (Darcy, Lorenz, Navier-Stokes, DED) shows the idea has breadth. The writing is clear, and the architecture is simple enough to reproduce in principle.\n\nThe soft spots, in order of severity. First, the input asymmetry above. Second, the OOD claims overreach: the paper says 'zero-shot generalizability to unseen initial conditions,' but the Lorenz experiment keeps initial conditions fixed and only varies parameters and loadings; no experiment uses unseen initial conditions. Third, the interpretability results are post-hoc alignments to formulas derived from the same equations the model was trained on; that's consistency checking, not physics discovery. The BDF5 coefficients in Eq. (19) aren't standard BDF5, and the 'closely matches' is qualitative. Minor: no error bars, no code/data release, and the Lorenz window arithmetic is confusingly stated (996 windows, then 995 after discarding a partial window).\n\nMy take: the architecture deserves a serious look, but the empirical comparison as written is unfair and the claims go beyond what is tested. A revised version with baselines given the same inputs, an ablation without F_{m+1}, and honest OOD language could be a solid paper. As is, I wouldn't cite it for performance numbers, but I'd consider citing the architectural idea if it survives a fair test. Yes, send to review — the flaw is fixable and the idea is substantive.","headline":"A promising architectural idea undermined by an unfair comparison: ASNO sees future forcing while baselines do not.","tokens_in":15557,"tokens_out":3461,"would_cite":false,"duration_ms":36934,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65L06"],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting temporal extrapolation from spatial correction yields a neural surrogate that generalizes zero-shot to unseen physics.","keywords":["spatio-temporal neural operator","backward differentiation formula","implicit-explicit scheme","nonlocal attention operator","zero-shot generalization","scientific machine learning","operator learning","Darcy flow"],"falsifier":"On the Darcy benchmark, test ASNO with the true next-step forcing $F_{m+1}$ withheld — set it to zero or to the previous step's forcing $F_m$ while keeping the same past five states — and compare the loss jump to the OOD-f shift; if the error rises to or beyond the baseline models' levels, the reported zero-shot performance depends on the model being handed the future forcing and the claim collapses to interpolation of a known input. Separately, recompute the learned NAO kernel on permeability microstructures drawn from a different GRF family and check it against $K_{\\text{true}}$ of Eq. (22): the interpretability claim fails if the kernel agrees only on training-distribution data.","tokens_in":14497,"feed_emoji":"⚛️","tokens_out":10384,"duration_ms":100374,"temperature":0.7,"pith_summary":"The paper sets out to show that a surrogate model for evolving physical systems predicts better, generalizes to unseen conditions, and reveals its own mechanics when its architecture mirrors a classical numerical scheme instead of treating prediction as an undifferentiated black box. The scheme in question is the implicit-explicit backward differentiation formula: an explicit step extrapolates the next state from past states, and an implicit step corrects that extrapolation using external forcing and spatial coupling. ASNO instantiates this split with a Transformer encoder for the explicit extrapolation and a nonlocal attention operator that learns a spatial kernel for the correction, yielding the composition $X_{m+1} = \\text{NAO}(\\text{TE}(X_{m-4},\\dots,X_m), F_{m+1})$. On Darcy flow, the Lorenz system, Navier–Stokes, and a directed-energy-deposition manufacturing dataset, the paper reports the lowest test losses, the lowest out-of-distribution losses, and slower error growth in long autoregressive rollouts than the compared baselines. A sympathetic reader would care because zero-shot adaptation to new physical parameters, and components that can be checked against known physics, are exactly what surrogate models for digital twins and engineering design currently lack.","feed_headline":"Time-space split beats physics baselines with zero-shot gains","feed_subtitle":"The learned pieces match a classical multistep scheme and the true Darcy kernel, not just the loss curves.","key_machinery":"The load-bearing object is the ASNO composition $X_{m+1} = \\text{NAO}(\\text{TE}(X_{m-4},\\dots,X_m), F_{m+1})$, which encodes the implicit-explicit BDF split. In the explicit step the Transformer encoder compresses the past five states into a latent $H_{m+1}$ that plays the role of the homogeneous extrapolation $\\tilde{X}_{m+1}$ in BDF. In the implicit step the Nonlocal Attention Operator (NAO), an attention mechanism that estimates a data-dependent nonlocal kernel $K[H_{1:d}, F_{1:d}](y,z)$ and outputs $X_{m+1}(y) = \\int K(y,z) F_{m+1}(z)\\,dz$, plays the role of the static PDE solver that corrects the extrapolation under forcing. The kernel map is built from stacked attention layers with residual connections (Eqs. 12–14), and the whole pipeline is trained end-to-end on the $L^2$ prediction loss over all systems (Eq. 16). The separation is what does the work: each module has a reduced job, one learning a multistep linear recurrence and the other a compact spatial correction, and the paper credits this split for both the generalization and the interpretability results.","core_discovery":"The paper's central claim is that spatio-temporal physics admits a separable decomposition that can be learned directly from data: the next state is a linear multistep extrapolation of recent history, corrected by a nonlocal operation that depends on that extrapolation and on the external loading at the next step. ASNO learns both pieces — a Transformer encoder produces the latent extrapolation $H_{m+1}$, and the Nonlocal Attention Operator (NAO) learns a kernel $K[H_{1:d}, F_{1:d}](y,z)$ and applies it to the forcing field to produce the output — and the paper reports that the jointly trained components rediscover known structure. On dynamic Darcy flow, the learned latent closely matches the BDF5 extrapolation (Eq. 19) and the learned kernel agrees with the closed-form kernel $K_{\\text{true}} = -(60/137)\\Delta t (I + (60/137)\\Delta t A)^{-1}$ from the Darcy discretization (Eq. 22). Empirically the paper finds that ASNO achieves the lowest best, OOD-forcing, and OOD-permeability test losses among FNO, U-Net, Transolver, GNOT, DeepONet, and Transformer variants, and that this advantage persists on the Lorenz system, the Navier–Stokes benchmark, and a DED melt-pool temperature task. The intended consequence is that a single trained model can predict unseen initial conditions, loadings, and environments zero-shot while exposing temporal and spatial contributions separately for physics discovery.","pith_inferences":["A direct test of the temporal component would extract the learned attention weights from the Transformer encoder and compare them coefficient-by-coefficient against the BDF5 coefficients of Eq. (19); the paper shows visual agreement on Darcy only, and a coefficient-level match would confirm genuine integration-scheme discovery.","The requirement that $F_{m+1}$ be known at inference marks a real boundary of the method: in forecasting settings where future forcing is itself unknown, ASNO would need a forcing forecaster, and the paper does not test that regime; a stress test with lagged or noised forcing would map that boundary.","The OOD-f experiment shifts forcing amplitude by a factor of 200 in one jump; sweeping the amplitude continuously between the training value and the OOD value would reveal whether generalization degrades gradually or snaps at a threshold, which is the practically useful curve for engineering deployment.","The kernel-versus-theory comparison protocol could be repeated on systems with analytically known kernels, such as the heat or wave equation, to sharpen the physics-discovery claim beyond the single Darcy example."],"forward_implications":["On dynamic Darcy flow the model posts the lowest best, OOD-forcing, and OOD-permeability losses (0.0368, 0.0673, 0.0982), beating the strongest baseline Transolver (0.0428, 0.0721, 0.1535) with fewer trainable parameters.","The learned components are inspectable: the Transformer latent tracks the BDF5 extrapolation and the NAO kernel matches the closed-form Darcy kernel, so the model can be read as having rediscovered the time integrator and the kernel structure of the PDE.","Long-term autoregressive rollouts stay stable, with cumulative error growing more slowly than for the baselines on Darcy and Navier–Stokes, which matters when predictions are fed back as inputs.","The same architecture transfers to an engineering use: DED melt pool temperature fields are predicted with 2.50% MAPE from laser power, scan rate, and laser location inputs.","On the Lorenz system the separation of a linear multistep forecast from a small nonlinear correction keeps chaotic trajectories near the true attractor longer than monolithic models, with test loss 0.000794 versus 0.000835 for the next best model."],"supporting_citations":[{"why":"Supplies the Nonlocal Attention Operator (NAO), whose attention-built kernel map and linear-activation recipe ASNO uses for the implicit spatial correction step.","marker":"(Yu et al., 2024)"},{"why":"Provides the implicit-explicit (IMEX) two-phase decomposition that motivates separating the explicit temporal extrapolation from the implicit spatial correction in Eq. (17).","marker":"(Ascher et al., 1995)"},{"why":"Supplies the FNO baseline and the 2D Navier–Stokes benchmark setup (varying viscosity over 100 profiles at 30×30 resolution) that ASNO is evaluated on.","marker":"(Li et al., 2020)"},{"why":"Supplies the DeepONet baseline for operator learning that ASNO must beat on Darcy flow, Lorenz, and Navier–Stokes.","marker":"(Lu et al., 2021)"},{"why":"Supplies the GNOT baseline, a general neural operator transformer compared across all PDE benchmarks.","marker":"(Hao et al., 2023)"},{"why":"Supplies the Transolver baseline, the strongest competitor on Darcy and Navier–Stokes, which ASNO is claimed to outperform.","marker":"(Wu et al., 2024)"},{"why":"Provides the self-attention transformer architecture from which the temporal encoder component is built.","marker":"(Vaswani et al., 2017)"},{"why":"Provides the GAMMA GPU-accelerated FEA solver that generates the DED melt pool temperature training data for the engineering case study.","marker":"(Liao et al., 2023)"},{"why":"Supplies the U-Net time-series baseline used in the Darcy and Navier–Stokes comparisons.","marker":"(Zhang et al., 2018)"}],"fun_headline_variants":["ASNO splits time and space, beats physics baselines zero-shot","Neural operator decouples time and space, then wins on physics","Time-space split operator: zero-shot gains on unseen physics","ASNO learns BDF5 and Darcy kernel from data alone","Physics discovery: ASNO beats baselines with zero-shot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the next state truly decomposes as a linear multistep extrapolation of past states corrected by a function that depends only on that extrapolation and on the next-step forcing (Eq. 17), an inductive bias imposed by the architecture rather than derived from the dynamics, and one that requires the future forcing $F_{m+1}$ to be known at prediction time.","fun_headline_variants_meta":{"raw":{"variants":["ASNO splits time and space, beats physics baselines zero-shot","Neural operator decouples time and space, then wins on physics","Time-space split operator: zero-shot gains on unseen physics","ASNO learns BDF5 and Darcy kernel from data alone","Physics discovery: ASNO beats baselines with zero-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2882,"prompt_tokens":1068,"completion_tokens":1814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":1727}},"tokens_in":684,"tokens_out":1814,"duration_ms":19485,"temperature":1.0,"reasoning_tokens":1727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:10:52.321303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the Darcy benchmark, test ASNO with the true next-step forcing $F_{m+1}$ withheld — set it to zero or to the previous step's forcing $F_m$ while keeping the same past five states — and compare the loss jump to the OOD-f shift; if the error rises to or beyond the baseline models' levels, the reported zero-shot performance depends on the model being handed the future forcing and the claim collapses to interpolation of a known input. Separately, recompute the learned NAO kernel on permeability microstructures drawn from a different GRF family and check it against $K_{\\text{true}}$ of Eq. (22): the interpretability claim fails if the kernel agrees only on training-distribution data.","supporting_citations":[{"cited_title":"M., Ruuth, S","cited_arxiv_id":null,"evidence_quote":"Provides the implicit-explicit (IMEX) two-phase decomposition that motivates separating the explicit temporal extrapolation from the implicit spatial correction in Eq. (17)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GNOT baseline, a general neural operator transformer compared across all PDE benchmarks."},{"cited_title":"N., Kaiser, L., and Polosukhin, I","cited_arxiv_id":null,"evidence_quote":"Provides the self-attention transformer architecture from which the temporal encoder component is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GAMMA GPU-accelerated FEA solver that generates the DED melt pool temperature training data for the engineering case study."}],"review_version":1}