{"id":"5986cb10-aa23-4d58-9e4e-080dc0e25a53","arxiv_id":"2608.04708","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Across four PDE benchmarks, transfer learning from low-fidelity weights is the only multi-fidelity neural operator strategy that consistently outperforms the high-fidelity-only baseline, while direct LF-input methods degrade as the fidelity gap grows.","lead":"This paper compares four ways to combine cheap low-resolution simulations with a few expensive high-resolution ones when training neural-network PDE surrogates. It finds that transfer learning, initializing the high-fidelity model with low-fidelity weights, is the only strategy that consistently beats the high-fidelity-only baseline, while directly feeding low-fidelity outputs into the model can amplify errors when the two fidelity levels disagree strongly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The modified-Darcy test case relies on a rank-2 POD projection of the HF snapshots (Eqs. 19-20); if the SVD includes test data, LF test inputs leak HF information, and even without leakage the projection cannot create true model-form discrepancies like RANS vs LES.","rationale":"The reader's weakest assumption correctly identifies that the POD-based LF data in Section 3.3 is a deterministic projection of the HF data and therefore cannot emulate structurally different model-form discrepancies. I sharpen that concern into a concrete leakage risk: since the paper never states whether the SVD in Eq. 19 is computed over the full snapshot matrix or only the training split, the benchmark may be giving the LF test inputs access to HF test information. That would be a correctness issue rather than merely a representativeness issue, and it would weaken the third test case's ability to support the paper's central claim about transfer learning under large model-form mismatch. At the same time, I do not think this concern overturns the reader's CONDITIONAL verdict: the smoke inflow case and the two grid-based cases still provide independent evidence that transfer learning is robust in the configurations tested, and the paper already discloses several limitations. The main adjustment I would ask for is a clear statement and re-run of the POD split, plus a genuinely model-form-different LF benchmark before generalizing to RANS-versus-LES scenarios. I therefore keep the verdict unchanged: CONDITIONAL, with the condition being that the modified-Darcy benchmark must be shown to be free of test leakage and, ideally, replaced or supplemented by a true model-form discrepancy case.","tokens_in":20899,"tokens_out":4620,"duration_ms":57496,"concrete_test":"Inspect the data-generation code or script for Section 3.3 to determine whether the SVD in Eq. 19 is fit on all snapshots or only on training snapshots. If it is fit on all snapshots, rerun the modified-Darcy experiments with POD modes computed exclusively from HF training snapshots, then project test snapshots with those fixed modes; if the transfer-learning margin over the HF baseline disappears or shrinks materially, the reported ranking is contaminated by test leakage. Independently, generate an alternative LF dataset that genuinely changes the PDE model, such as a coarse-mesh solution of a different permeability model or a simplified governing equation, at a comparable discrepancy, and verify whether transfer learning still consistently outperforms the other strategies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.3, the paper replaces mesh coarsening with the rank-2 POD projection A2 = U Sigma_2 V^T (Eqs. 19-20) and claims this 'mimics' governing-equation differences. Two problems follow. First, the text does not state whether the SVD in Eq. 19 is computed on the full snapshot matrix A, including the test split, or only on training snapshots. If the full matrix is used, the low-fidelity representation of each test input is built from the HF test solution's own dominant modes, so the benchmark leaks test information through the LF channel; that alone would invalidate the third test case as a ranking experiment. Second, even with a correct train/test split, a rank-2 projection of HF solutions only removes fine detail. The LF and HF fields share the same POD basis by construction, so the discrepancy is an information-loss gap, not a model-form gap of the kind the paper motivates with RANS versus LES. The central claim that transfer learning is robust under 'substantial model discrepancies' is therefore not established for genuine model-form mismatch; it is established only for projection-type large discrepancies. This matters because the two introduced benchmarks are presented as the paper's main contribution, and the transfer-learning-only conclusion relies on the modified Darcy results as the key large-discrepancy case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates four multi-fidelity neural operator strategies—intermediate, two-step, residual, and transfer learning—against a high-fidelity-only baseline across four PDE test cases: a one-dimensional stochastic Poisson equation, a two-dimensional Darcy flow, a modified Darcy flow intended to emulate governing-equation discrepancies, and an unsteady smoke inflow problem. The main empirical claim is that only transfer learning consistently outperforms the high-fidelity baseline across all scenarios, while methods that feed low-fidelity predictions into the high-fidelity model are prone to error amplification under large LF-HF discrepancy, especially in autoregressive time-stepping. The paper also introduces two new test cases: a POD-truncated Darcy flow and a time-dependent smoke inflow benchmark.","tokens_in":21051,"tokens_out":4175,"duration_ms":47158,"significance":"If the central empirical claims hold, the paper is a useful systematic comparison for the multi-fidelity neural operator community. Its strengths include a consistent evaluation protocol with repeated random train/test splits, a clear comparison against a high-fidelity-only baseline, and a new unsteady smoke inflow dataset that addresses an underrepresented temporal setting. The observation that transfer learning is more robust than direct LF-input injection under large discrepancies is practically relevant and falsifiable. However, the modified Darcy test case, which is the key large-discrepancy case supporting the transfer-learning conclusion, rests on a construction that is not a faithful emulation of model-form discrepancies, and the text does not rule out test information leakage through the SVD. These issues are load-bearing for the paper's main claim and require careful resolution.","major_comments":[{"comment":"The paper does not state whether the SVD in Eq. (19) is computed on the full snapshot matrix A, including the test split, or only on training snapshots. If the full matrix is used, the low-fidelity representation of each test input is constructed from the dominant modes of the high-fidelity test solutions themselves, which leaks information from the test set into the inputs. This would invalidate the third test case as a ranking experiment. The manuscript must specify the split used and, if the SVD included test data, rerun the experiment with a training-only SVD.","section":"Section 3.3, Eqs. (19)-(20)"},{"comment":"Even with a correct train/test split, a rank-2 POD projection of the HF solution does not emulate a governing-equation discrepancy such as RANS versus LES. The LF and HF fields share the same POD basis by construction, so the LF field is a projection of the HF field that only removes spatial detail. This is an information-loss gap, not a model-form gap, and it cannot produce the physically different solutions that a genuinely different model would yield. Consequently, the conclusion that transfer learning is robust under 'substantial model discrepancies' is not established for true model-form mismatches; it is established only for projection-type discrepancies. The authors should either construct the LF data from an independent solver or model, or substantially weaken the claim that this test case resembles governing-equation differences.","section":"Section 3.3, Eqs. (19)-(20) and Fig. 10"},{"comment":"The text states that the training rollout is restricted to a single step ('we restrict the rollout to a single step', and 'the HF model is trained with a rollout number of one'), but Table 6 lists Rollout number 3 for both LF and HF models. This contradiction matters because the paper attributes the two-step model's failure to error accumulation in an autoregressive setting. If the HF model was actually trained with rollout 3, the stated training protocol is wrong; if it was trained with rollout 1, the table is wrong and the claims about autoregressive error accumulation should be reconsidered or clarified.","section":"Section 3.4 and Table 6"}],"minor_comments":[{"comment":"The first numeric cell in the table reads '504.43×10−2', which appears to be a typographical error; the intended value is likely '4.43×10−2'. The row labels for the first two columns are also difficult to parse.","section":"Table 2"},{"comment":"The text says 'the FNO yields a better performance compared to other kernels, such as the Wavelet neural operator (WNO), in this test case' but the comparison is not shown. This claim should either be supported with a displayed result or removed.","section":"Section 3.1"},{"comment":"Section 3.2 states that both transfer learning and the two-step architecture are the only approaches that consistently surpass the HF baseline across all training sample configurations in that test case, while the abstract and conclusion state that only transfer learning consistently outperforms the baseline across all scenarios. The scope of the 'only transfer learning' claim should be clarified, since two-step also achieves this in the 2D Darcy case.","section":"Section 3.2 and Abstract"},{"comment":"The claim that transfer learning is 'the least sensitive' to increasing discrepancy is based on only two discrepancy levels per model. The figure shows two points per model, so the slope comparison is not statistically supported; this should be acknowledged or supplemented with more discrepancy levels.","section":"Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The paper's strengths are real, but the modified Darcy construction is central to the 'only transfer learning' conclusion, and the current text leaves open both a test leakage risk and a mismatch between the claimed model-form discrepancy and what a POD projection actually provides. These issues are fixable either by clarifying and re-running with a train-only SVD, or by reframing the claims as being about projection-type large discrepancies. I would also ask the editor to encourage the authors to release code or detailed data-generation scripts, since the benchmark value of the two new test cases depends on reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful benchmark paper for people working on multi-fidelity neural operators. The four-strategy comparison (two-step, residual, intermediate, transfer learning) across four test cases is the most systematic I've seen in this subfield, and the internal ranking is clear: transfer learning beats the HF-only baseline in every configuration, while methods that inject LF predictions into the HF input degrade when the LF-HF gap is large, especially in autoregressive time-stepping. The two new test cases—POD-truncated Darcy and unsteady smoke inflow—are a real step beyond the usual grid-coarsening benchmarks. The tables and figures are consistent with the stated ranking, and the paper is honest about several limitations.\n\nThe soft spots are also real, and one of them is load-bearing. The modified Darcy case is built by projecting the HF solution snapshots onto two POD modes. The text never says whether that SVD is computed on the training split only or on the full dataset. If it includes test data, the LF representation of each test input is built from its own dominant modes—that's leakage and the case can't rank anything. If the split is clean, you still haven't created a model-form gap like RANS vs LES; you've created an information-loss gap where LF and HF share the same basis by construction. The abstract and conclusion frame the transfer-learning winner as robust under 'substantial model discrepancies,' and that framing is only as strong as this proxy. I'd want the authors to either compute the SVD on training snapshots, state it explicitly, and reframe the case as 'large projection-based discrepancy,' or replace it with a genuinely different solver.\n\nThe smaller issues are addressable: alpha in Eq. 14 is never reported; Table 6 lists rollout 3 while Section 3.4 says rollout 1; Fig. 12's sensitivity claim rests on two points per model; no code is released. None of these kill the paper, but together they make the experiments hard to reproduce. The citation pattern is fine—prior multi-fidelity works get proper credit.\n\nVerdict: this deserves a serious referee and a major revision. The benchmark cases and the transfer-learning observation are worth keeping; the POD case needs to be fixed or honestly reframed, and the reporting gaps need to be closed.","headline":"Useful benchmark comparison with a solid internal ranking, but the 'model-form fidelity' claim rests on a POD-projection proxy that needs to be fixed or reframed before the paper is publishable.","tokens_in":21746,"tokens_out":3015,"would_cite":true,"duration_ms":31706,"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":"Transfer learning is the only multi-fidelity strategy that consistently beats the high-fidelity baseline across four PDE benchmarks; direct low-fidelity-input injection degrades as the fidelity gap grows.","keywords":["neural operator","multi-fidelity learning","transfer learning","Fourier neural operator","proper orthogonal decomposition","Darcy flow","smoke inflow","autoregressive error accumulation"],"falsifier":"Build a version of the Darcy or smoke-inflow benchmark where the low-fidelity data comes from a solver with a genuinely different governing equation or closure (for example, an actual turbulence model rather than a finer or coarser grid), rerun all four strategies, and check whether transfer learning still wins; if two-step or residual matches it there, the central claim is limited to the paper's projection-style fidelity gap.","tokens_in":20557,"feed_emoji":"💨","tokens_out":10410,"duration_ms":101305,"temperature":0.7,"pith_summary":"Multi-fidelity neural operators try to compensate for scarce high-fidelity (HF) simulation data by exploiting abundant low-fidelity (LF) data. This paper systematically compares four ways to do that — two-step, residual, intermediate, and transfer learning — on four PDE test cases, including two newly designed to create large LF-HF gaps: a Darcy flow whose LF solution is obtained by keeping only the two dominant proper orthogonal decomposition modes, and an unsteady smoke-inflow problem whose LF-HF discrepancy grows over time. Its central finding is that only transfer learning, which initialises the HF neural operator with weights pretrained on LF data, consistently outperforms a model trained on HF data alone. Methods that feed LF predictions into the HF model as inputs tend to amplify errors, most severely in autoregressive time-stepping (using earlier predictions as inputs to later steps), when the LF and HF solutions disagree substantially. The intended upshot is that practitioners should prefer transfer learning for multi-fidelity PDE surrogate modelling when fidelity levels differ by more than grid resolution.","feed_headline":"Only transfer learning beats the high-fidelity baseline in every test","feed_subtitle":"When low- and high-fidelity solutions diverge strongly, only weight-transfer from low-fidelity data helps; injecting LF predictions hurts.","key_machinery":"The machinery is a controlled comparison of four multi-fidelity integration schemes on one baseline architecture (Fourier neural operator, with a wavelet operator in the Darcy cases), plus two purpose-built benchmarks that separate fidelity-gap size from grid effects. The modified Darcy case creates its low-fidelity data by rank-2 proper orthogonal decomposition, $A_2=U\\Sigma_2 V^T$ (Eqs. 19–20), raising the mean absolute LF-HF discrepancy from 0.02647 to 0.151; the smoke-inflow case pairs a 96×120 low-fidelity grid with a 256×320 high-fidelity grid in a time-dependent smoke-advection problem, where discrepancy grows with time and single-step rollout during training is followed by autoregressive evaluation. These benchmarks are what let the paper attribute performance differences to the LF-HF mismatch mechanism rather than to trivial resolution similarity.","core_discovery":"The paper's central claim is that the way low-fidelity information enters a neural operator determines whether multi-fidelity training helps or hurts. On all four test cases — 1D stochastic Poisson, 2D Darcy flow, a modified Darcy flow with a large fidelity gap, and unsteady smoke inflow — the transfer-learning strategy is the only one that consistently beats the HF-only baseline. The two-step and residual strategies, which form $\\hat{y}_{HF}=F(x,\\hat{y}_{LF})$ or learn the residual $r=y_{HF}-y_{LF}$, and the intermediate architecture, which shares a trunk and emits LF from a hidden layer, all lose accuracy when the LF-HF discrepancy is large; in the unsteady case the gap widens over time and autoregressive rollout compounds the error. The paper interprets this as evidence that injecting LF outputs as input features confuses the HF network, whereas LF-pretrained weights supply a structural prior that HF fine-tuning can refine.","pith_inferences":["Editorial inference: the POD-truncated Darcy 'model-form' gap is not a true governing-equation difference — the LF data is a rank-2 projection of the HF snapshots, so it shares the HF mode structure and can only remove detail; the transfer-learning conclusion is therefore not established for genuinely different physics until a test with LF from a different solver is run.","Editorial inference: the discrepancy-error analysis suggests a practical diagnostic — measure the LF-HF gap before choosing a strategy, using transfer learning when the gap is large and any of the direct-injection methods when it is small.","Editorial inference: the mechanism at work, that weight-space transfer tolerates strong fidelity mismatch while input-space injection does not, is not specific to neural operators and could be tested in simpler multi-fidelity regression settings where the same failure mode should appear."],"forward_implications":["With scarce high-fidelity data, initialising the HF neural operator with low-fidelity-pretrained weights is the reliable way to exploit LF data; it improves accuracy over HF-only training on all tested problems.","Two-step, residual, and intermediate architectures that inject LF predictions into the HF input should be avoided when the LF-HF gap is large, since they can underperform the HF baseline.","In time-dependent problems, direct LF injection is especially risky: the fidelity gap widens with time and autoregressive rollout turns small errors into large ones.","The benefit of LF data shrinks as HF data becomes plentiful, and non-transfer strategies can slightly degrade performance once 50 HF samples are available.","Benchmark design matters: close-to-HF low-fidelity data (grid-only differences) can mask differences between multi-fidelity strategies that appear only at larger fidelity gaps."],"supporting_citations":[{"why":"Supplies the original multi-fidelity wavelet neural operator benchmarks and the residual architecture that the comparison extends.","marker":"[31]"},{"why":"Introduces transfer learning with a Fourier neural operator for multi-fidelity fluid flow, the central strategy being tested.","marker":"[29]"},{"why":"Provides another transfer-learning multi-fidelity FNO formulation that supports the paper's chosen strategy.","marker":"[30]"},{"why":"Defines the intermediate and multi-step (two-step) neural network architectures that the paper adapts to neural operators.","marker":"[27]"},{"why":"Supplies the nonlinear correction formulation for combining LF and HF data that underlies the direct-injection approaches.","marker":"[26]"},{"why":"Source of the 2D Darcy flow problem and the fair-comparison methodology used to design the benchmark cases.","marker":"[47]"},{"why":"Introduces the unsteady smoke-inflow dataset used as the new time-dependent multi-fidelity test case.","marker":"[48]"},{"why":"Provides the proper orthogonal decomposition used to construct the modified Darcy low-fidelity data.","marker":"[7]"}],"fun_headline_variants":["Only transfer learning beats HF baseline on hard PDE cases","Fidelity gap: transfer learning thrives where feeding LF fails","Rather than two-step, transfer learning wins when fidelity gap is large","Transfer learning beats HF baseline when LF-HF gap is wide"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's third test case assumes that filtering the high-fidelity solutions down to their two dominant patterns mimics the gap between two genuinely different physics models, even though the low-fidelity data is derived from the very high-fidelity solutions it is compared against, so the transfer-learning conclusion may not hold for true model-form discrepancies.","fun_headline_variants_meta":{"raw":{"variants":["Only transfer learning beats HF baseline on hard PDE cases","Fidelity gap: transfer learning thrives where feeding LF fails","Rather than two-step, transfer learning wins when fidelity gap is large","Transfer learning beats HF baseline when LF-HF gap is wide"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000475,"raw_usage":{"total_tokens":2383,"prompt_tokens":996,"completion_tokens":1387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":1318}},"tokens_in":612,"tokens_out":1387,"duration_ms":11972,"temperature":1.0,"reasoning_tokens":1318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:26:59.020668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a version of the Darcy or smoke-inflow benchmark where the low-fidelity data comes from a solver with a genuinely different governing equation or closure (for example, an actual turbulence model rather than a finer or coarser grid), rerun all four strategies, and check whether transfer learning still wins; if two-step or residual matches it there, the central claim is limited to the paper's projection-style fidelity gap.","supporting_citations":[{"cited_title":"Tripura, A","cited_arxiv_id":null,"evidence_quote":"Supplies the original multi-fidelity wavelet neural operator benchmarks and the residual architecture that the comparison extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces transfer learning with a Fourier neural operator for multi-fidelity fluid flow, the central strategy being tested."},{"cited_title":"Multi-fidelity Fourier Neural Operator for Fast Modeling of Large-Scale Geological Carbon Storage","cited_arxiv_id":"2308.09113","evidence_quote":"Provides another transfer-learning multi-fidelity FNO formulation that supports the paper's chosen strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the intermediate and multi-step (two-step) neural network architectures that the paper adapts to neural operators."},{"cited_title":"Zakaria, A","cited_arxiv_id":null,"evidence_quote":"Introduces the unsteady smoke-inflow dataset used as the new time-dependent multi-fidelity test case."},{"cited_title":"Chatterjee, An introduction to the proper orthogonal decomposition, Current Science 78 (7) (2000) 808–817","cited_arxiv_id":null,"evidence_quote":"Provides the proper orthogonal decomposition used to construct the modified Darcy low-fidelity data."}],"review_version":1}