{"id":"c2a8da85-09ed-4704-aca1-f343678c7c5e","arxiv_id":"2602.00072","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A surjective normalizing flow, pretrained on low-fidelity data and fine-tuned on high-fidelity data, produces probabilistic structural-response surrogates that match high-fidelity accuracy with far fewer high-fidelity simulations.","lead":"This paper proposes a two-stage generative surrogate: a normalizing flow is first trained on cheap low-fidelity simulations, then fine-tuned on a small set of expensive high-fidelity simulations to give probabilistic predictions with uncertainty. If it works as shown, engineers could get high-fidelity accuracy from mostly cheap data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Predictive uncertainty is never validated: no calibration/coverage check for the claimed 95% credibility intervals; the probabilistic claim is empirically unsupported.","rationale":"The reader's weakest assumption about latent-space capacity is plausible, but I find a more direct and more easily settled gap: the paper never checks whether its predictive distributions are calibrated. The reader already assigned CONDITIONAL based on missing evidence, so the verdict need not change; adding a calibration analysis is a further condition. The math of the surjective flow follows the SurVAE framework and does not appear internally inconsistent, but the empirical support for the central probabilistic claim is incomplete without coverage evaluation. I therefore keep the conditional status.","tokens_in":14229,"tokens_out":6946,"duration_ms":80011,"concrete_test":"For all 20 held-out HF test inputs in both Case 1 and Case 2, compute the empirical coverage of the 95% credibility intervals at each time step and over whole trajectories, and construct a calibration curve (e.g., expected calibration error or quantile-quantile plot). If the empirical coverage deviates substantially from the nominal 95%, the uncertainty quantification claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central value proposition is probabilistic prediction: the Introduction and Sec. 4.4 argue that unlike deterministic TCN, the proposed model 'produces full predictive probability distributions rather than point estimates,' and the Conclusion claims 'reliable uncertainty estimates.' However, no calibration or coverage check is reported anywhere. In Secs. 4.1–4.2, 95% credibility intervals are shown visually for only two test pairs per case, and Sec. 4.4 reports only mean relative L2 errors. The generative decoder of Eq. (16) is a diagonal Gaussian over the discarded 190-dimensional block, imposing strong conditional independence given the 10-dimensional latent. If these predictive intervals are miscalibrated, the claim of 'probabilistic predictions with HF accuracy' fails even if predictive means are close. The manuscript provides no evidence to rule this out.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage probabilistic multi-fidelity surrogate modeling framework. A conditional surjective normalizing flow is first pretrained on abundant low-fidelity (LF) simulation data and then fine-tuned on a small high-fidelity (HF) dataset. To handle high-dimensional time-series responses, the architecture uses a funnel layer that reduces the state from 200 to 10 dimensions while retaining exact log-likelihood training; a diagonal Gaussian decoder models the discarded block. The approach is validated on a reinforced-concrete slab finite-element model under two scenarios (strong and weak LF-HF correlation), including a comparison against a multi-fidelity temporal convolutional network (TCN) baseline. The paper claims HF-accurate probabilistic predictions with reliable uncertainty estimates and a significant improvement over LF-only baselines while using fewer HF evaluations.","tokens_in":14441,"tokens_out":8456,"duration_ms":88858,"significance":"The methodological idea is timely and practically relevant: if validated, it would allow engineers to replace many expensive HF simulations with cheap LF simulations plus a small HF fine-tuning set, while obtaining calibrated probabilistic predictions. The derivation of the surjective-layer likelihood in Sec. 3.2 is explicit and internally consistent, and the benchmark is a nontrivial structural dynamics application. The inclusion of a deterministic TCN baseline is a useful sanity check. However, the empirical evidence presented is not sufficient to establish the central probabilistic claims: no calibration/coverage validation is reported, and the key accuracy comparisons are single-point estimates without statistical spread. The paper would be significantly stronger with repeated runs, uncertainty calibration checks, and sensitivity studies of the core architectural choices.","major_comments":[{"comment":"The paper's central value proposition is probabilistic prediction, yet no calibration or coverage check is reported. The 95% CIs are shown only for two test pairs per case (Figs. 5, 6, 9, 10), and the number of Monte Carlo samples used to compute means and CIs is not stated. Since the decoder in Eq. (16) is a diagonal Gaussian on a 190-dimensional discarded block, the intervals could be systematically miscalibrated even if predictive means are accurate. Please add empirical coverage on the 20 held-out pairs (e.g., fraction of true responses inside the 95% CI, PIT histograms) and report the MC sample count. Without this, the conclusion of 'reliable uncertainty estimates' is unsupported.","section":"Secs. 4.1–4.2, 5"},{"comment":"The headline comparison rests on a single point estimate: average relative L2 error 0.091 for TCN vs 0.0791 for the proposed method over 20 test datasets. No variance, confidence interval, paired significance test, or repeated runs with different seeds are reported, and TCN hyperparameters are fixed without tuning. Moreover, the abstract claims 'significantly outperforms LF-only baselines,' but no systematic LF-only baseline is included in the error comparisons; only two examples show LF R^2 values. Please add distributions over repeated runs and a statistical test, and clarify what 'fewer HF evaluations' is compared with.","section":"Sec. 4.4, 4.3"},{"comment":"The architecture fixes the latent dimension Q=10 and layer dimensionalities [200,200,200,200,10,10,10] without any sensitivity study. The capacity of the surjective layer and the Gaussian decoder to represent the LF-HF discrepancy depends critically on Q and decoder size. Please include ablations varying Q (e.g., 5, 20, 50) or at least a discussion of how Q was selected. As it stands, the reported HF accuracy may be configuration-specific rather than a general property of the framework.","section":"Sec. 4.1, Eq. (16)"}],"minor_comments":[{"comment":"There are two abstracts in the manuscript. The first claims validation on 'a rail-sleeper-ballast and a reinforced concrete slab,' while the full-text abstract and the body present only the reinforced-concrete slab benchmark. This inconsistency must be resolved.","section":"Abstract"},{"comment":"Equation (8) is written as a pointwise equality, but the right-hand side depends on a sample z ~ q(z|y,theta). In the SurVAE framework this is an expectation over q or a lower-bound expression. Please clarify the notation so that the subsequent derivation is unambiguous.","section":"Sec. 3.2, Eq. (8)"},{"comment":"The heading 'Bijecive conditional NFs' contains a typo; it should be 'Bijective'. Also, Eq. (21) contains a duplicated word ('provides provides').","section":"Sec. 3.1"},{"comment":"The reported 12.7% reduction appears inconsistent with the stated values 0.091 vs 0.0791, which corresponds to approximately 13.1%. Please verify the arithmetic.","section":"Sec. 4.4"},{"comment":"The HF-only model is described as 'training a single NF model using only 180 HF data pairs,' but its architecture, training epochs, and batch size are not specified. State whether it uses the same surjective architecture and training schedule as the proposed method for a fair comparison.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core is plausible, but the empirical evaluation currently does not support the probabilistic claims. The missing calibration analysis and lack of statistical rigor are substantial but fixable within the scope of the manuscript. I would not accept in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward combination of published pieces: surjective normalizing flows (SurVAE/Funnels) as the conditional density model, and transfer learning from low-fidelity to high-fidelity data. The math in Sec. 3 is internally consistent; the likelihood decomposition is a restatement of Nielsen et al. and Klein et al. The new part is the application: using this to build multi-fidelity surrogates for a reinforced concrete slab benchmark, with different LF-HF correlation regimes. That is a legitimate new application, and the framework is plausible.\n\nWhat the paper does well: it is clearly written, the model description is precise, and the authors are honest that the components are not new. The empirical demonstration, while limited, does show the expected behavior: fine-tuned model beats LF-only and HF-only baselines on the few test cases shown. The relative L2 comparison against TCN on 20 test sets is a step in the right direction.\n\nWhere it falls short: the evidence is far too thin for the claims. The strongest quantitative results rest on two selected test pairs per case. There is no repeated training, no sensitivity analysis for the latent dimension Q or architecture, and no error bars around the aggregate L2 numbers. More importantly, the paper's central value proposition is probabilistic prediction, but predictive intervals are never validated. No calibration or coverage check appears anywhere; the 95% CIs are shown for two examples and interpreted visually. For a method whose selling point is uncertainty quantification, that is a load-bearing gap. The decoder in Eq. (16) is a diagonal Gaussian over the discarded 190-dimensional block, so conditional independence is assumed; without calibration evidence, the claim of 'reliable uncertainty estimates' is unsupported.\n\nThere is also a mismatch: the abstract promises a rail-sleeper-ballast benchmark, but the paper only presents the RC slab. That needs reconciliation.\n\nThe math is not the problem. The evaluation is. If the authors release code/data, run more repetitions, report aggregate statistics with variability, add a sensitivity study for Q and architecture, and actually check whether the claimed intervals have the right coverage, this could be a serviceable application paper. As it stands, it is a promising but unproven demonstration.","headline":"A coherent application of SurVAE/Funnels to multi-fidelity surrogate modeling, but the empirical case is thin and the uncertainty claims are unvalidated.","tokens_in":14920,"tokens_out":1664,"would_cite":false,"duration_ms":16831,"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":"A two-stage normalizing-flow surrogate aims to match high-fidelity accuracy using mostly cheap low-fidelity simulations.","keywords":["multi-fidelity modeling","surrogate modeling","normalizing flows","transfer learning","uncertainty quantification","surjective layers","structural dynamics","generative models"],"falsifier":"Concrete check: hold out high-fidelity responses from a time segment where the signal is sharply transient, fine-tune the flow on the remaining responses, and count how often the claimed 95% credibility interval actually contains the simulated high-fidelity response; if coverage falls far below 95% in those segments, or if varying the latent dimension from small to large never reproduces the reported accuracy, the central claim is called into question.","tokens_in":14114,"feed_emoji":"🏗️","tokens_out":7168,"duration_ms":73652,"temperature":0.7,"pith_summary":"The paper tries to solve a practical data-scarcity problem: high-fidelity simulations are too expensive to generate in the quantities deep-learning surrogates need, while low-fidelity simulations are cheap but biased. Its proposal is a two-stage training scheme: first pretrain a normalizing-flow generative model on abundant low-fidelity data, then fine-tune it on a small high-fidelity dataset so the flow learns to correct the low-fidelity bias. The central mechanism is a surjective (dimension-reducing) flow layer that compresses high-dimensional time-series responses into a 10-dimensional latent code while preserving exact-likelihood training. If the proposal works, engineers could substitute mostly cheap coarse-mesh simulations plus a small fine-mesh set for a large fine-mesh dataset, and still obtain probabilistic predictions with quantified uncertainty rather than point estimates. The paper demonstrates this on a reinforced-concrete slab benchmark under both strong and weak low-fidelity/high-fidelity correlation, reporting high-fidelity-level accuracy in both regimes.","feed_headline":"Cheap simulations plus small fine-tune match high-fidelity accuracy","feed_subtitle":"Pretrained on many cheap simulations and fine-tuned on few costly ones, it returns uncertainty-aware predictions.","key_machinery":"Central object: a surjective normalizing flow, a neural density model that maps data to a simple base distribution through invertible layers with exact likelihoods. Since standard flows preserve dimension, the paper inserts a dimension-reducing 'funnel' layer that splits each response into kept and discarded blocks, maps the kept block to a 10-dimensional latent code with a coupling transform, and models the discarded block with a Gaussian decoder. The exact log-likelihood becomes base log-density plus decoder log-density plus a Jacobian correction. This compression makes large-scale low-fidelity pretraining and small high-fidelity fine-tuning tractable, and fine-tuning transfers knowledge b","core_discovery":"The paper claims that a surjective normalizing flow, pretrained on abundant low-fidelity data and then fine-tuned on a small high-fidelity dataset, produces probabilistic predictions with high-fidelity accuracy while using far fewer high-fidelity simulations. The key architectural idea is a dimension-reducing 'funnel' layer that compresses 200-step response time series to a 10-dimensional latent code while keeping the exact log-likelihood computable. In the reinforced-concrete slab benchmark, the fine-tuned model tracks the high-fidelity response closely with narrow credibility intervals, even under weak low-fidelity/high-fidelity correlation, and the paper reports that it beats a determinis","pith_inferences":["A natural next test the paper does not run: measuring calibration of the 95% credibility intervals (does the true high-fidelity response fall inside them about 95% of the time?), which would tell whether the uncertainty estimates are trustworthy, not just narrow.","The same two-stage scheme could likely be coupled with active learning—choosing which high-fidelity simulations to run next—to reduce the required fine-tuning set further; the paper does not explore this.","The 10-dimensional latent code may itself encode physical structure (e.g., dominant modes of the response); inspecting it could give interpretability and a diagnostic for when transfer is working.","One could test the framework when low- and high-fidelity models differ by simplified physics rather than mesh refinement, a harder and arguably more realistic transfer scenario than the excitation perturbation used in the paper's Case 2."],"forward_implications":["If the central claim holds, expensive fine-mesh simulations can be largely replaced by coarse-mesh simulations plus a modest fine-mesh set, cutting the cost of building probabilistic surrogates for structural dynamics.","The same pretrain-then-fine-tune recipe should apply to other high-dimensional responses, such as fields or long time series, wherever standard bijective flows are impractical due to their dimension-preserving constraint.","Downstream tasks that need uncertainty—reliability analysis, Bayesian updating, risk-informed decisions—can use full predictive distributions from the surrogate instead of point estimates.","Because the paper's ablation shows that pretraining on low-fidelity data helps even when high-fidelity data is scarce, the approach offers a concrete answer to the data-hunger of deep surrogates."],"fun_headline_variants":["Cheap sims pretrain, costly sims fine-tune: surrogate hits HF accuracy","Few costly runs fine-tune a flow pretrained on many cheap runs","Transfer learning cuts high-fidelity data need for probabilistic surrogates","Surrogate pretrains on low-fidelity, fine-tunes on few high-fidelity cases","Generative flow + transfer learning: high-fidelity accuracy from few costly runs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The approach stands on the assumption that a 10-dimensional latent code, plus a Gaussian decoder for the discarded part of the response, is expressive enough to represent the difference between low- and high-fidelity behavior for unseen inputs; the paper fixes this capacity without a sensitivity study.","fun_headline_variants_meta":{"raw":{"variants":["Cheap sims pretrain, costly sims fine-tune: surrogate hits HF accuracy","Few costly runs fine-tune a flow pretrained on many cheap runs","Transfer learning cuts high-fidelity data need for probabilistic surrogates","Surrogate pretrains on low-fidelity, fine-tunes on few high-fidelity cases","Generative flow + transfer learning: high-fidelity accuracy from few costly runs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001276,"raw_usage":{"total_tokens":5085,"prompt_tokens":802,"completion_tokens":4283,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":4193}},"tokens_in":546,"tokens_out":4283,"duration_ms":30566,"temperature":1.0,"reasoning_tokens":4193,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:11:04.092701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Concrete check: hold out high-fidelity responses from a time segment where the signal is sharply transient, fine-tune the flow on the remaining responses, and count how often the claimed 95% credibility interval actually contains the simulated high-fidelity response; if coverage falls far below 95% in those segments, or if varying the latent dimension from small to large never reproduces the reported accuracy, the central claim is called into question.","supporting_citations":[],"review_version":1}