{"id":"6dede4a3-a9cf-426c-acaa-68ea9fa76c34","arxiv_id":"2412.19510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LoRA fine-tuning of a pretrained InversionNet model on OpenFWI matches full fine-tuning in-distribution and improves out-of-distribution generalization for seismic full-waveform inversion.","lead":"Researchers tested whether a pretrained neural network for seismic imaging (trained on the public OpenFWI dataset) can be fine-tuned cheaply with LoRA, a parameter-efficient method, to adapt to new geological scenarios. They report that LoRA matches full fine-tuning in-distribution and often beats it on out-of-distribution tests, while updating only 1.1 million of 24.4 million parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD superiority of LoRA-PFM over FFT-PFM is not statistically established: Table IV contains counterexamples and no error bars or training-budget controls are reported.","rationale":"After reading the paper and the reader's verdict, I agree that the work has a plausible recipe and the released code is a plus. However, the central claim of OOD superiority is the load-bearing assertion, and the evidence for it is fragile. The reader focused on the synthetic-to-field gap, which is an external validity concern; I focus on the internal validity of the reported numbers. The internal contradictions and data errors mean the quantitative support for the claimed OOD benefit is unreliable. A well-controlled re-run with seeds and significance testing would settle whether the effect is real. This is a concrete, feasible check. If the effect does not survive, the paper's contribution reduces to a standard LoRA-adaptation recipe without demonstrated OOD benefit. If it does survive, the paper is a useful empirical study. Therefore the conditional verdict is appropriate, and the proposed test is the minimal further evidence required.","tokens_in":18403,"tokens_out":8668,"duration_ms":71947,"concrete_test":"Re-run the Section IV-E OOD protocol with 5 random seeds per method and identical compute budgets: same number of epochs, same batch size, and the same learning-rate schedule for Baseline, FFT-PFM, and LoRA-PFM, with validation-based early stopping allowed for all methods. Record mean and standard deviation of MAE/RMSE/SSIM for all 12 train-test pairs, and compute a paired Wilcoxon signed-rank test across all metric-task cells comparing LoRA-PFM vs FFT-PFM. Also recompute Table VI to check for transcription errors. If LoRA-PFM does not win a statistically significant majority of cells, the paper's headline claim of OOD superiority is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-E claims that 'LoRA-PFM performs better than Baseline and FFT-PFM' across OOD datasets, but Table IV shows the opposite in multiple cells: trained on Style B, tested on CurveFault B, FFT-PFM has MAE 0.255 vs LoRA-PFM 0.276 and SSIM 0.452 vs 0.410; trained on Style A, tested on Style B, FFT-PFM has MAE 0.085 vs 0.088. The text itself acknowledges 'FFT-PFM slightly outperforms LoRA-PFM on FlatVel B and CurveFault B' after claiming universal LoRA superiority. Section IV-D contains a direct contradiction: 'LoRA-PFM performs better than FFT-PFM in all the tested datasets' despite Table III showing FFT-PFM with lower MAE/RMSE on FVB and STA. No error bars, seeds, or significance tests are reported, and fine-tuning hyperparameters (epochs, learning rate, early stopping) are not specified separately for FFT-PFM vs LoRA-PFM. The supplementary low-data tables also contain impossible entries (Table VI, 75%: RMSE 0.178 < MAE 0.312), indicating data-entry errors. Therefore, the central OOD advantage could be a consequence of noise, undertraining of FFT-PFM, or tabulation error rather than a genuine PEFT property.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pretrained foundational model (PFM) for seismic full-waveform inversion (FWI), built by supervised pretraining an InversionNet backbone on six 2D OpenFWI datasets. It then compares two adaptation strategies on four downstream FWI tasks: full fine-tuning of all 24.4M parameters (FFT-PFM) and low-rank adaptation (LoRA-PFM) that updates only 1.1M parameters. The authors claim that FFT-PFM outperforms task-specific baselines, that LoRA-PFM achieves comparable in-distribution performance to FFT-PFM while reducing memory, and that LoRA-PFM outperforms both FFT-PFM and baselines on out-of-distribution (OOD) tasks and in low-data regimes. The paper also reports memory-efficiency and storage benefits of LoRA adapters.","tokens_in":18627,"tokens_out":3834,"duration_ms":33980,"significance":"If substantiated, the result would be practically valuable: it would show that parameter-efficient fine-tuning is sufficient to adapt a pretrained FWI model to new geological scenarios, reducing per-task storage and compute. The paper has clear strengths: it uses the public OpenFWI benchmark, releases code, reports parameter counts, and includes a systematic low-data study. However, the central OOD-superiority claim is currently not supported by the paper's own numbers, and several tables contain internal inconsistencies. The contribution is incremental (LoRA applied to InversionNet) but the study could be useful if the claims are corrected and the evidence is strengthened with statistical rigor.","major_comments":[{"comment":"The text states that 'LoRA-PFM performs better than FFT-PFM in all the tested datasets,' but Table III shows the opposite for three of the four datasets: on FVB, FFT-PFM has MAE 0.030 vs LoRA-PFM 0.031, RMSE 0.080 vs 0.084, and SSIM 0.952 vs 0.943; on STA, FFT-PFM has better MAE, RMSE, and SSIM; on STB, FFT-PFM again has better metrics. Only on CFB do the MAE values tie (0.139). This direct contradiction must be resolved, either by correcting the claim or by presenting additional evidence that the differences are not meaningful.","section":"Section IV-D, Table III"},{"comment":"The section claims that 'LoRA-PFM performs better than the Baseline and FFT-PFM' across OOD datasets and later concludes that 'for OOD samples, LoRA-PFM performs better than FFT-PFM across all datasets.' Table IV contains multiple counterexamples: when trained on Style B and tested on CurveFault B, FFT-PFM achieves MAE 0.255 vs LoRA-PFM 0.276, RMSE 0.339 vs 0.391, and SSIM 0.452 vs 0.410; when trained on Style A and tested on Style B, FFT-PFM has MAE 0.085 vs 0.088 and SSIM 0.680 vs 0.667. The text itself acknowledges that 'FFT-PFM slightly outperforms LoRA-PFM on FlatVel B and CurveFault B' in the Style-B-trained row, which is inconsistent with the universal conclusion. The OOD claim needs to be restricted to the datasets where it actually holds, or supported by a statistical test that accounts for these exceptions.","section":"Section IV-E, Table IV"},{"comment":"No error bars, random seeds, or significance tests are reported for any of the ID or OOD comparisons, and the fine-tuning hyperparameters (number of epochs, learning rate, batch size, early stopping criteria) are not specified separately for FFT-PFM and LoRA-PFM. The supplementary gives pretraining details (e.g., 120 epochs, AdamW, WarmupMultiStepLR) but does not state whether the two fine-tuning methods were given identical training budgets or how convergence was determined. Without this information, the reported OOD advantages of LoRA-PFM over FFT-PFM could be an artifact of undertrained FFT-PFM or of a single favorable seed. The paper should report multi-seed means and standard deviations and explicitly state the compute budget allocated to each method.","section":"Section IV-E and Section IV-A"},{"comment":"Supplementary Table VI contains an impossible entry: for test CurveFault B at 75% training data, FFT-PFM reports RMSE 0.178 with MAE 0.312, which violates the mathematical relationship RMSE ≥ MAE. The same table also shows identical rows for 10% and 25% training data (FFT-PFM 0.277/0.372/0.465 and LoRA-PFM 0.259/0.353/0.504), suggesting a tabulation or copy error. These issues undermine confidence in the low-data quantitative results and require the tables to be regenerated and rechecked.","section":"Supplementary Table VI"}],"minor_comments":[{"comment":"The text says that LoRA-PFM uses 1.1 million parameters, 'which is a reduction of 4.5%.' This is numerically incorrect: 1.1M is approximately 4.5% of 24.4M, so the statement should say 'a reduction of 95.5%' or 'using only 4.5% of the parameters.'","section":"Section IV-G"},{"comment":"The sentence 'highlighting its robustness as a preferred technique for diverse applications' contains a typo: 'ODD datasets' should be 'OOD datasets.'","section":"Section IV-E"},{"comment":"The model is described as 'task-agnostic,' but it is pretrained via supervised learning on six FWI datasets to predict velocity maps. Consider replacing 'task-agnostic' with 'multi-task pretrained' or clarifying the intended meaning, since the pretraining task is not agnostic to the FWI objective.","section":"Abstract and Section III-B"},{"comment":"The hyperparameter description is ambiguous: it states 'convergence was achieved after 90 epochs' and later 'the models undergo training for 120 epochs.' Please clarify which of these applies to pretraining and which, if any, applies to fine-tuning, and provide fine-tuning-specific hyperparameters for both FFT-PFM and LoRA-PFM.","section":"Supplementary Section VII"},{"comment":"The caption says the model was tested with 'FlatVel B (blue), and Style A (red),' but panel (c) shows Style B. The caption should list all three test sets (FlatVel B, Style A, Style B) as shown in the figure.","section":"Figure 6 caption"},{"comment":"In the row for train=FVB, test=STA, the method entry reads 'FFT-PFM M'; the stray 'M' should be removed.","section":"Table IV"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a plausible and potentially useful empirical study, but the central claims are overstated relative to the evidence in the tables. The internal contradictions between the text and Table III/Table IV, combined with the absence of error bars and the impossible entries in the supplementary tables, suggest that the manuscript needs a careful revision of both the claims and the reported numbers. I recommend asking the authors to correct the tables, rephrase the universal claims, and add multi-seed statistics and fine-tuning budget details before resubmission. The work is within scope for the journal as an applied PEFT study, but the current version does not yet meet the standard of evidence for its headline conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read. The useful core: pretraining InversionNet on six OpenFWI 2D datasets, then fine-tuning with LoRA (1.1M trainable params) gives results broadly on par with full fine-tuning on the same benchmarks, and sometimes better on cross-dataset OOD tests, at much lower memory cost. That's a practical, legitimate empirical contribution to the DL-FWI subfield. It's not a new method—LoRA and transfer learning are established—but the application is real, and the code is released.\n\nThe paper's central weakness is that its claims outrun its tables. Section IV-D states LoRA-PFM 'performs better than FFT-PFM in all the tested datasets,' but Table III shows FFT-PFM with lower MAE/RMSE on FlatVel B and Style A. Section IV-E first claims universal OOD superiority, then later concedes FFT-PFM slightly wins on FlatVel B and CurveFault B when trained on Style B. These aren't stylistic nits; they're the paper's headline. Also, there are no error bars, no multiple seeds, no significance tests, and fine-tuning hyperparameters aren't reported separately for the two methods. So the OOD advantage could be noise or undertraining.\n\nThe supplementary tables contain clear data errors. Table VI has identical rows for 10% and 25% training on FlatVel B, and at 75% the FFT-PFM RMSE (0.178) is lower than its MAE (0.312) on the same data—impossible. That makes the low-data conclusions difficult to trust without a re-tabulation.\n\nWhat holds up? The ID results in Table III show LoRA-PFM within a hair of FFT-PFM, which is a useful finding by itself. The broad approach—pretrained InversionNet plus PEFT—is credible and likely of practical value for geophysics practitioners. The flaws are fixable, and the central direction isn't invalidated.\n\nWho is this for? Applied geophysicists and DL-FWI researchers wanting a practical PEFT comparison. It deserves serious peer review—the topic is relevant and the code is public—but only as a conditional. My recommendation: engage, ask for a corrected analysis with seeds and error bars, and make the claims match the tables. Worth a reading group if you want to talk about how overclaiming creeps into otherwise useful empirical papers.","headline":"Useful PEFT-for-FWI recipe undermined by overclaimed superiority and data errors in the tables; fixable, worth a conditional review.","tokens_in":19198,"tokens_out":3912,"would_cite":false,"duration_ms":35384,"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 pretrained seismic inversion model, adapted by low-rank updates to just 1.1 million of its 24.4 million parameters, matches full fine-tuning on familiar geological data and beats it on unseen geological styles.","keywords":["full-waveform inversion","parameter-efficient fine-tuning","Low-Rank Adaptation","foundation model","seismic velocity estimation","out-of-distribution generalization","transfer learning","OpenFWI benchmark"],"falsifier":"Run the same LoRA-PFM versus full-fine-tuning comparison on real or realistic field-seismic data with different acquisition geometries, noise levels, and missing traces; if LoRA-PFM's out-of-distribution advantage does not appear there, or if the two methods tie within measurement noise, the central out-of-distribution claim fails. A cheaper check is to repeat the out-of-distribution evaluation across all 12 OpenFWI dataset families as train-test pairs and verify that LoRA-PFM beats full fine-tuning on the majority of pairs rather than only on the four reported here.","tokens_in":18173,"feed_emoji":"🌊","tokens_out":6209,"duration_ms":56071,"temperature":0.7,"pith_summary":"The paper claims that seismic full-waveform inversion (FWI), which recovers subsurface velocity maps from recorded seismic wavefields, is better done by pretraining one general model on diverse geological structures and then adapting it rather than training a fresh model for each geological scenario. The pretrained foundational model, built on the InversionNet encoder-decoder architecture and six two-dimensional OpenFWI datasets spanning flat, curved, and faulted velocity structures, outperforms task-specific baselines when fully fine-tuned. The central comparison is between full fine-tuning of this pretrained model and Low-Rank Adaptation (LoRA), a parameter-efficient method that trains only 1.1 million adapter parameters instead of all 24.4 million. LoRA matches full fine-tuning on in-distribution tasks, beats it on out-of-distribution datasets, and keeps that advantage when only 10 to 100 percent of the training data is available. The authors state that their experiments are confined to synthetic OpenFWI simulations, so the real-field value of the claim is an open question that the paper acknowledges.","feed_headline":"Updating 1.1M of 24.4M parameters matches full fine-tuning","feed_subtitle":"Seismic velocity inversion gets a pretrained base model whose lightweight adapters generalize better and use less memory.","key_machinery":"The load-bearing object is the Low-Rank Adaptation (LoRA) module: for each convolutional and transposed-convolutional weight matrix $W_0$, the update is $W = W_0 + \\alpha B A$, with trainable low-rank matrices $A$ and $B$, a common rank $r=16$, and scaling $\\alpha=16$, while the pretrained weights stay frozen. These low-rank updates are inserted into the InversionNet U-shaped encoder-decoder, giving roughly 1.1 million trainable parameters out of 24.4 million. The pretrained base model supplies general geological features, and the small adapters supply task-specific adjustments, which is why the authors can swap adapters across tasks instead of storing full copies of the model. This design is what carries the paper's efficiency and generalization claims.","core_discovery":"On the paper's own terms, the discovery is that a foundation-model strategy transfers to a geophysical imaging task: pretraining InversionNet on a mixture of six OpenFWI datasets produces a feature set that improves downstream velocity recovery over training from scratch, and the cheapest way to exploit it is LoRA rather than full fine-tuning. With rank 16 and alpha 16, LoRA-PFM updates only 1.1 million parameters, yet its test metrics stay within a small margin of full fine-tuning on the four target datasets (for example, MAE 0.031 versus 0.030 on FlatVel B and 0.139 versus 0.139 on CurveFault B). When models trained on one dataset are tested on other dataset families with different geological structure, LoRA-PFM consistently beats full fine-tuning in MAE, RMSE, and SSIM, and the margin grows when only 10 to 25 percent of the training data is available. The authors interpret this as evidence that freezing the pretrained weights and nudging them with low-rank residuals preserves general seismic features while preventing the overfitting that full fine-tuning suffers on scarce or shifted data.","pith_inferences":["If the LoRA advantage generalizes to field data, a practical consequence is that a seismic contractor could maintain one frozen subsurface model and ship tiny per-survey adapters, making multi-client model storage and transfer far cheaper.","The same recipe, pretrain broadly and add LoRA adapters, could be tested on other geophysical inverse problems such as electrical resistivity tomography or ground-penetrating radar inversion, where task-specific networks show similar out-of-distribution failures.","Comparing LoRA against adapter or prompt-tuning variants on the same pretrained FWI model would separate the benefit of low-rank updates from the benefit of parameter efficiency in general.","An ablation that trains LoRA on an untrained random-weight InversionNet and compares it with LoRA on the pretrained base would quantify how much of the out-of-distribution gain comes from pretraining rather than from low-rank adaptation itself."],"forward_implications":["Pretraining one InversionNet-style model on multiple geological families, then fully fine-tuning it, is a stronger recipe than training a task-specific model from scratch on the same target dataset.","LoRA adaptation of the pretrained model gives near-identical in-distribution accuracy while training only 1.1 million of 24.4 million parameters, so adaptation is cheaper in memory and storage.","On held-out geological families, LoRA-PFM improves MAE, RMSE, and SSIM relative to full fine-tuning, so adapter-based fine-tuning is the recommended route for out-of-distribution deployment.","The LoRA advantage persists when the fine-tuning set is cut to 10 percent of the training data, meaning the approach is suited to low-data adaptation.","Because the frozen base model stays intact, one pretrained model can serve many tasks by swapping small adapters rather than storing a full model copy per task."],"supporting_citations":[{"why":"Supplies the OpenFWI dataset families and splits used for pretraining and for in-distribution and out-of-distribution evaluation.","marker":"[3]"},{"why":"Defines InversionNet, the encoder-decoder architecture that serves as both the pretrained foundational model's backbone and the task-specific baseline.","marker":"[5]"},{"why":"Introduces Low-Rank Adaptation (LoRA), the parameter-efficient fine-tuning mechanism whose rank-16 adapters carry the paper's efficiency and out-of-distribution claims.","marker":"[15]"},{"why":"Provides the PEFT taxonomy and the general argument that updating a fraction of parameters reduces cost while preserving transfer performance.","marker":"[14]"}],"fun_headline_variants":["Seismic foundation model: 1.1M-parameter LoRA beats full tuning on unseen data","LoRA on a pretrained seismic model matches full fine-tuning with 95% fewer updates","Foundation model for seismic FWI: small fine-tune beats training from scratch","Pretraining on diverse seismic data makes LoRA the best adapter for new geology","Seismic inversion: 1.1M tuned parameters beat full retraining on unseen data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's load-bearing premise is that the distribution shifts among the OpenFWI dataset families stand in for real geological change, and the authors explicitly note that their own experiments are confined to synthetic simulations.","fun_headline_variants_meta":{"raw":{"variants":["Seismic foundation model: 1.1M-parameter LoRA beats full tuning on unseen data","LoRA on a pretrained seismic model matches full fine-tuning with 95% fewer updates","Foundation model for seismic FWI: small fine-tune beats training from scratch","Pretraining on diverse seismic data makes LoRA the best adapter for new geology","Seismic inversion: 1.1M tuned parameters beat full retraining on unseen data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001659,"raw_usage":{"total_tokens":6585,"prompt_tokens":944,"completion_tokens":5641,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":5529}},"tokens_in":560,"tokens_out":5641,"duration_ms":37127,"temperature":1.0,"reasoning_tokens":5529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:15:55.625925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same LoRA-PFM versus full-fine-tuning comparison on real or realistic field-seismic data with different acquisition geometries, noise levels, and missing traces; if LoRA-PFM's out-of-distribution advantage does not appear there, or if the two methods tie within measurement noise, the central out-of-distribution claim fails. A cheaper check is to repeat the out-of-distribution evaluation across all 12 OpenFWI dataset families as train-test pairs and verify that LoRA-PFM beats full fine-tuning on the majority of pairs rather than only on the four reported here.","supporting_citations":[{"cited_title":"Openfwi: Large-scale multi-structural benchmark datasets for full waveform inversion,","cited_arxiv_id":null,"evidence_quote":"Supplies the OpenFWI dataset families and splits used for pretraining and for in-distribution and out-of-distribution evaluation."},{"cited_title":"Inversionnet: An efficient and accurate data-driven full waveform inversion,","cited_arxiv_id":null,"evidence_quote":"Defines InversionNet, the encoder-decoder architecture that serves as both the pretrained foundational model's backbone and the task-specific baseline."}],"review_version":1}