{"id":"dffd38d7-cc7a-41d3-bb6f-5dfd0a4757ce","arxiv_id":"2506.14784","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A ConvNet pretrained on CFD surface pressures can be retrained with only its last layers to predict angle of attack and onflow speed in a new domain or for a new task, matching offline accuracy in the studied cases.","lead":"This paper tests whether a neural network, pretrained on simulated airfoil pressure data, can be quickly updated to predict angle of attack and flow speed under a new flow model, a wider operating range, or a new prediction task. The updates worked in the studied cases, but the same approach did not help when the pressure readings were noisy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim lacks a from-scratch-on-target baseline: Fig. 12 shows transfer learning reaches source-domain accuracy, but without retraining on the target domain we cannot tell whether frozen features provide any benefit over ordinary retraining.","rationale":"The paper is a competent, honestly reported engineering demonstration, and the reader's CONDITIONAL verdict is reasonable. My stress-test focuses on one gap that is more than a completeness issue: the absence of a from-scratch-on-target baseline is the operational test of the paper's core transferability assumption. The central claim in Section IV.C is that transfer learning with frozen convolutional layers recovers offline-level accuracy on the target domain. But the only target-domain model shown is N_TL; there is no model trained directly on D_R from scratch. If such a model matches N_TL, then the result is explained by the target data alone, not by knowledge transfer, and the phrase \"potential of the approach\" is unsupported for the headline domain-adaptation scenario. The domain-extension demonstration has the same structural gap. The task-adaptation case happens to include a true from-scratch baseline (N_OL(τ_V∞)), and there the one-layer transfer fails, which sharpens the concern: the method's benefit is not established where the proper control exists. I do not see this as grounds for rejection; the paper's claims are modest and its self-reported limitations (no real-time training, noisy-data failure, reduced training time) are plausible. But the missing control should be supplied, or the central claim should be weakened to say that transfer learning matches the source-trained network's accuracy on the target test set without claiming that the frozen features are the cause. In addition, all MAE values appear to be single runs without seed variance, so the proposed multi-seed comparison would also stabilize the conclusion. Therefore I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":20028,"tokens_out":4372,"duration_ms":51572,"concrete_test":"Retrain ConvNet-S and ConvNet-D from scratch on D_R (and on D_e for the extension case) using the same 72/18/10 split, min-max normalization, and the same optimizer/epochs/loss as the transfer-learning runs, for at least 5 random seeds. Compare test MAE for both α and V∞ against the N_TL values in Figs. 12 and 15. If from-scratch-on-target MAE is statistically indistinguishable from N_TL, then the paper's transferability claim is unsupported; if N_TL is significantly better, or matches from-scratch accuracy with materially fewer target samples, the claim is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.C's central result (Fig. 12) compares N_TL on the target domain D_R against N_OL on D_S and N_OL on D_R. The label \"N_OL on D_R\" is not a from-scratch target baseline; it is the source-trained network evaluated on target data. The proper control for the transferability premise of Section II — that convolutional features are generic and reusable — is a network trained from scratch on D_R with the same architecture, data split, and hyperparameters. Without that control, the observed recovery of source-level MAE on D_R could be explained by the target training set being sufficient to train the small unfrozen head, or even the whole network, without any benefit from the pretrained features. The domain-extension case (Fig. 15) has the same gap: \"N_OL on D_e\" is offline-on-D_i evaluated on D_e, not a from-scratch D_e model. Notably, in the one case where a true from-scratch target baseline is shown (Fig. 20, N_OL(τ_V∞)), the one-layer transfer fails and only after unfreezing 98.697% of the shallow network does it approach the baseline; this makes the missing baseline in the domain-adaptation claim more consequential, not less.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a transfer-learning framework for predicting two onflow parameters, angle of attack and freestream speed, from sparse surface-pressure data on an airfoil. A ConvNet is trained offline on a source domain, then its convolutional weights are frozen and selected fully connected layers are retrained on a target domain or for a changed prediction task. Demonstrations are carried out on steady RANS CFD data for the NLR 7301 airfoil, covering adaptation to a change in turbulence model (SA vs. RSM), extension of the operating domain, addition of sensor noise, and task transfer from angle-of-attack prediction to speed prediction. The paper reports that transfer learning recovers source-level accuracy in the domain-change case, improves extrapolation in the domain-extension case, and works for task adaptation only when two layers are retrained, while the noisy-data application is acknowledged to be ineffective.","tokens_in":20140,"tokens_out":6590,"duration_ms":76104,"significance":"If the transferability premise were established, the framework would be practically useful for updating onflow-parameter predictors when the data distribution shifts, with modest training-time savings relative to offline retraining (Table II). The paper has several strengths: it uses a held-out test-set protocol, reports training times with mean and standard deviation, clearly states the failure of the noise-adaptation scenario, and provides a useful negative result for one-layer task transfer. The central claim, however, is currently underdetermined because the key comparisons lack a from-scratch target-domain baseline, so the observed improvements could be attributable to ordinary retraining on the target data rather than to reuse of the pretrained convolutional features.","major_comments":[{"comment":"The central quantitative claim lacks the necessary control. The figure compares N_TL on the target domain D_R with N_OL on the source domain D_S, but the only target-domain comparator, 'N_OL on D_R', is the source-trained network evaluated on target data, not a network trained from scratch on D_R. Since Section II justifies freezing the convolutional layers by the premise that their features are 'generic and can be reused for target tasks', the paper must report a from-scratch target model trained with the same architecture, data split, and hyperparameters. Without such a baseline, the recovery of source-level MAE on D_R could be explained by the target training set being sufficient to train the small unfrozen head, or even the whole network, without any benefit from the pretrained features. Figure 20 shows that in the one scenario where a true from-scratch baseline is provided, the one-layer transfer fails and only after unfreezing 98.697% of the shallow network's weights does it approach the baseline; this makes the missing control in Fig. 12 consequential rather than pedantic.","section":"Section IV.C, Fig. 12"},{"comment":"The domain-extension claim has the same control gap. The label 'N_OL on D_e' denotes the source/initial-domain network evaluated on the extended domain, which is a zero-shot extrapolation result, not a from-scratch model trained on D_e. The improvement of N_TL on D_e over N_OL on D_e is expected simply because N_TL has been trained on D_e training data. To demonstrate that transfer from the initial domain is beneficial, the paper needs a from-scratch network trained on the merged D_i plus D_e data under the same protocol. Without this baseline, the reported improvement in MAE on the extended domain does not support the conclusion that transfer learning, rather than additional training data, is responsible.","section":"Section IV.D, Fig. 15"},{"comment":"The task-adaptation experiment itself undercuts the generality of the transferability premise. One-layer transfer produces no improvement over the source network, and the only successful configuration retrains 98.697% of the weights of the shallow architecture (Section IV.A). The statement in Section IV.F that 'transfer learning is not the same as nearly fully retraining the entire network' is technically true only because the convolutional layers remain frozen, but those layers contain a small fraction of the total weights; the model is in effect almost completely retrained. This should be framed as fine-tuning with most weights unfrozen, and the generic-feature claim should be tested by freezing the convolutional layers while retraining only the final layer(s) and comparing against a from-scratch target model.","section":"Section IV.F, Fig. 20 and Section IV.A"}],"minor_comments":[{"comment":"The MAE values in the bar charts are reported as single point estimates without training-seed variation; only Table II reports mean and standard deviation. Since the differences in Figs. 12 and 15 can be small on a log scale, it would be helpful to report repeated-run statistics or at least state how many seeds were used.","section":"Section IV.B, Figs. 7-12"},{"comment":"In the ConvNet-D parameter table, the row 'Conv4 (256, 256)' appears twice; if the second occurrence is meant to be Conv5, the table should be corrected.","section":"Section IV.A, Table III"},{"comment":"The appendix header says 'the parameters of the ConvNet-D, ConvNet-D, and FCNN architectures', where the second 'ConvNet-D' should presumably be 'ConvNet-S'.","section":"Appendix A.1"},{"comment":"All transfer-learning MAE demonstrations fix the dataset size at nd = 1024. Since the introduction motivates transfer learning for settings with limited or streaming data, the paper would be strengthened by varying the target-domain training set size in the transfer-learning phase.","section":"Sections IV.C-IV.F"},{"comment":"No code or data availability statement is included; providing the CFD datasets or trained models would improve reproducibility, given that all demonstrations are in-house generated.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable empirical study and the negative results (noise adaptation, one-layer task transfer) are honestly reported. However, the central claim about the benefit of transfer learning is not yet supported because the main figures lack a from-scratch target-domain baseline. The missing experiments are straightforward to run and would either confirm or refute the transferability premise, so major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a straightforward, honest application of standard transfer learning to a practical inverse problem: predicting angle of attack and speed from surface pressure on an airfoil. It does not introduce new methodology, but it does deliver a clean empirical comparison of frozen-feature retraining across four realistic scenarios. The authors also clearly report where the method fails, especially for noisy data and real-time training, which earns credit.\n\nWhat is actually new is the application: using a ConvNet pretrained on SA-turbulence-model CFD data, then retraining only the last one or two layers to handle RSM data, an extended alpha/V_inf domain, a noisy pressure domain, and a different output task. The offline phase shows ConvNets beating a fully connected net for both tasks, and the transfer results are internally consistent with the claim that freezing the convolutional layers works for related domain shifts. The task-adaptation case is particularly informative: one-layer transfer fails, two-layer transfer succeeds for the shallow net, and they honestly show that this requires unfreezing almost all weights.\n\nThe soft spots are real but not fatal. The stress-test note is correct: for the domain-adaptation and domain-extension cases there is no from-scratch-on-target baseline. N_OL on D_R is not that baseline; it is the source-trained model evaluated on target data. Without training a fresh model on D_R with the same architecture and data, you cannot tell whether the frozen features actually help or whether a simply retrained head (or full network) would do just as well. Given that the one case with a true baseline (task adaptation) shows transfer only matching the baseline after unfreezing 98.7% of the shallow net, this missing control matters. That said, the authors' stated claims are modest -- they claim \"potential\" and \"similar accuracy,\" not superiority over retraining -- so the gap weakens the contribution without collapsing it.\n\nOther issues: no code or data shipped, no variance across seeds for the central MAE numbers, and training hyperparameters are unreported. These make independent verification impossible, but the data-generation pipeline is described well enough that the results are plausible. The citation pattern is legitimate; the self-citation to the first author's earlier ConvNet airfoil work is a genuine base, not circularity.\n\nBottom line: this is a solid case study that deserves serious referee time, provided the authors add the from-scratch baseline for the domain-adaptation cases and report variance or seeds. It will be useful to people building pressure-based air-data estimators and to anyone comparing fine-tuning recipes for surrogate models. I would accept it for review, with the missing baseline as the main revision request.","headline":"A competent, honest engineering demonstration of transfer learning for pressure-based onflow estimation; the main gap is the missing from-scratch baseline, but the modest claims mostly hold up.","tokens_in":20827,"tokens_out":1768,"would_cite":false,"duration_ms":22320,"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 lets a convolutional network trained on CFD surface-pressure data predict angle of attack and onflow speed in a new flow domain by retraining only the final layers.","keywords":["transfer learning","convolutional neural network","surface pressure data","angle of attack prediction","onflow speed prediction","domain adaptation","task adaptation","RANS CFD"],"falsifier":"Run the same transfer-learning pipeline with target-domain data from a physical wind-tunnel experiment instead of a second CFD turbulence model: if the retrained network's test-set mean absolute error on the experimental pressure data does not approach the offline-trained baseline within the same order of magnitude, the claim that frozen convolutional features transfer to physically different pressure distributions is refuted.","tokens_in":19658,"feed_emoji":"🌌️","tokens_out":10772,"duration_ms":98991,"temperature":0.7,"pith_summary":"The paper proposes a transfer-learning recipe for predicting angle of attack and onflow speed from sparse surface-pressure measurements: train a convolutional neural network once on simulated pressure distributions, freeze its feature-extracting layers, and retrain only the final predictor layers on data from a new flow regime, a new task, or an extended envelope. The goal is to make data-driven onflow-parameter predictors adaptable with little new training, which matters for wind tunnel monitoring and for flight and wind turbine operations where sensor faults or changing conditions would break a fixed model. On CFD data for an NLR 7301 airfoil, the recipe recovers offline-level prediction accuracy when the turbulence model changes or the domain is extended, while changing the task from angle of attack to speed works only when the final two layers are retrained and adaptation to noisy pressure data is reported as not effective. If the recipe holds on real measurements, a reusable predictor could follow a changing data distribution at a fraction of the retraining cost.","feed_headline":"Airflow predictors adapt to new domains with frozen features","feed_subtitle":"Retraining just the final layers on new pressure data recovers offline-level accuracy for angle of attack and speed.","key_machinery":"The central mechanism is weight freezing with selective retraining. A ConvNet (the AlexNet-derived 1-D 'ConvNet-D' and the shallower 'ConvNet-S') is first trained offline on source-domain surface-pressure arrays; for transfer, a network with the same architecture is initialized from those weights and only the final fully connected layer or layers before the output are unfrozen, while all convolutional filters stay fixed. This operationalizes the assumption, taken from prior ConvNet transfer work, that the convolutional features are generic and reusable, so the regression head alone must adapt to the target domain or task. The trade study retrains either one layer (0.086% of ConvNet-S weights, 0.017% of ConvNet-D) or two layers (98.697% of ConvNet-S, 70.359% of ConvNet-D), and the outcome of task adaptation depends on retraining enough of that head.","core_discovery":"The paper's central claim is that a ConvNet trained to map 1-D arrays of airfoil surface pressure to onflow parameters can be adapted to a target domain with a different data distribution, to an extended domain, or to a new prediction task by freezing the weights of the convolutional layers and retraining only the fully connected layers immediately preceding the output node. The main evidence is that after transfer learning to a target domain generated with a different RANS (Reynolds-averaged Navier-Stokes) turbulence model (RSM instead of Spalart-Allmaras), the retrained network's test-set mean absolute error on the target domain is similar to what the offline-trained network achieved on the source domain, for both angle of attack and onflow speed. When the domain is extended beyond the offline-training envelope, transfer learning reduces the error in the extrapolated region, in the shallow architecture by up to an order of magnitude. Task adaptation from predicting angle of attack to predicting onflow speed succeeds for the shallow network only when two fully connected layers are retrained (98.697% of its weights, still with frozen convolutions), while the deeper network remains an order of magnitude worse than a dedicated offline-trained speed predictor. The paper also states that transfer learning to a noisy-pressure domain is not effective and that, although transfer learning shortens training time for larger datasets and denser architectures, real-time learning has not been achieved.","pith_inferences":["Editorial inference: if the frozen-feature assumption is as general as the paper suggests, the same recipe should transfer across airfoil geometries, not just across turbulence models; this can be tested by pretraining on one airfoil's pressure arrays and fine-tuning on another geometry's arrays.","Editorial inference: the paper's failure on noisy data implies the learned convolutional features are not robust to input corruption, so adding a denoising preprocessing step or unfreezing convolutional layers in the noise case is a natural next experiment the authors did not run.","Editorial inference: the real-world motivation depends on steady RANS CFD standing in for quasi-steady wind tunnel, flight, and wind turbine conditions, so a direct validation on experimental surface-pressure data is the critical step that would make the method operational."],"forward_implications":["A ConvNet trained on pressure data from one RANS turbulence model can be reused for a different turbulence model by retraining only the final predictor layers, reaching source-domain-level accuracy on the target domain.","An offline model trained on only part of the angle-of-attack and speed envelope can be extended to the full envelope by transfer learning, with errors in the extrapolated region falling by up to an order of magnitude for the shallow architecture.","Transferring the angle-of-attack predictor to the speed-prediction task works for the shallow network only when two fully connected layers are retrained, which unfreezes most of its weights while keeping the convolutions fixed.","Adapting to pressure data corrupted by Gaussian noise does not transfer well: the retrained network's errors stay much higher than in the other demonstration cases, so noise robustness requires a different approach.","Transfer learning reduces training time relative to offline retraining, particularly for larger datasets and denser networks, but the resulting training times still exceed the limits for real-time learning."],"supporting_citations":[{"why":"Supplies the transfer-learning premise that convolutional features are generic and reusable, which justifies freezing the feature extractor.","marker":"[38]"},{"why":"Supports the claim that deeper convolutional feature hierarchies learn more discriminative, transferable features.","marker":"[51]"},{"why":"Provides the domain/task taxonomy (inductive, transductive, unsupervised) used to frame the demonstration cases.","marker":"[49]"},{"why":"The DLR-TAU flow solver used to generate all steady RANS source and target pressure datasets.","marker":"[60]"},{"why":"Spalart-Allmaras one-equation turbulence model defines the source domain data distribution D_S.","marker":"[61]"},{"why":"SSG/LRR-lnω Reynolds stress model defines the target domain data distribution D_R in the main domain-change demonstration.","marker":"[62]"},{"why":"AlexNet is converted to 1-D form to create the deeper ConvNet-D architecture.","marker":"[57]"},{"why":"DLR SMARTy toolbox supplies the design-of-experiments capability used to sample angle of attack and speed.","marker":"[63]"},{"why":"NLR 7301 supercritical airfoil is the geometry whose surface-pressure arrays serve as network inputs.","marker":"[59]"},{"why":"Halton sequences provide the low-discrepancy, sequentially extendable sampling used to build the CFD datasets.","marker":"[64]"}],"fun_headline_variants":["Frozen features let airflow predictors adapt to new domains","Retraining only final layers adapts airflow prediction across domains","Transfer learning with frozen features handles domain and task shifts","Airflow prediction adapts to new tasks by retuning just the top layers","Pressure-based airflow model transfers to new domains with minimal retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach rests on the premise that the convolutional features learned from source-domain pressure distributions are generic enough to be reused for target domains and tasks, so freezing all but the final layers costs little accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Frozen features let airflow predictors adapt to new domains","Retraining only final layers adapts airflow prediction across domains","Transfer learning with frozen features handles domain and task shifts","Airflow prediction adapts to new tasks by retuning just the top layers","Pressure-based airflow model transfers to new domains with minimal retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1705,"prompt_tokens":1103,"completion_tokens":602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":518}},"tokens_in":719,"tokens_out":602,"duration_ms":6292,"temperature":1.0,"reasoning_tokens":518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:03:08.517538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same transfer-learning pipeline with target-domain data from a physical wind-tunnel experiment instead of a second CFD turbulence model: if the retrained network's test-set mean absolute error on the experimental pressure data does not approach the offline-trained baseline within the same order of magnitude, the claim that frozen convolutional features transfer to physically different pressure distributions is refuted.","supporting_citations":[{"cited_title":"Oquab, L","cited_arxiv_id":null,"evidence_quote":"Supplies the transfer-learning premise that convolutional features are generic and reusable, which justifies freezing the feature extractor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the claim that deeper convolutional feature hierarchies learn more discriminative, transferable features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the domain/task taxonomy (inductive, transductive, unsupervised) used to frame the demonstration cases."},{"cited_title":"Schwamborn, T","cited_arxiv_id":null,"evidence_quote":"The DLR-TAU flow solver used to generate all steady RANS source and target pressure datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Spalart-Allmaras one-equation turbulence model defines the source domain data distribution D_S."},{"cited_title":"Eisfeld, C","cited_arxiv_id":null,"evidence_quote":"SSG/LRR-lnω Reynolds stress model defines the target domain data distribution D_R in the main domain-change demonstration."},{"cited_title":"Krizhevsky, I","cited_arxiv_id":null,"evidence_quote":"AlexNet is converted to 1-D form to create the deeper ConvNet-D architecture."},{"cited_title":"Bekemeyer, A","cited_arxiv_id":null,"evidence_quote":"DLR SMARTy toolbox supplies the design-of-experiments capability used to sample angle of attack and speed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NLR 7301 supercritical airfoil is the geometry whose surface-pressure arrays serve as network inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Halton sequences provide the low-discrepancy, sequentially extendable sampling used to build the CFD datasets."}],"review_version":1}