{"id":"ab5ed381-dca6-4ecf-aa96-a88208a9a73f","arxiv_id":"2502.01111","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A class-conditioned diffusion model, pre-trained on synthetic data and iteratively fine-tuned with self-generated labels on field data, performs four seismic processing tasks with one network.","lead":"Researchers built a single generative diffusion model, GSFM, trained on synthetic seismic data and then fine-tuned on unlabeled field data, that handles denoising, backscattered noise attenuation, interpolation, and low-frequency extrapolation. If it holds up, seismic processing could move from many task-specific networks to one adaptable foundation model with built-in uncertainty estimates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Field-data claims rest on pseudo-labels generated by the model itself; reported improvements may be self-convergence, not correctness.","rationale":"The reader's weakest_assumption correctly identifies the load-bearing weakness: the field-data pipeline and its quantitative evaluation rely on the model's own pseudo-labels as ground truth. I agree with this concern. Synthetic experiments provide genuine supporting evidence for the x0-targeted diffusion architecture: GSFM consistently beats Benchmark 1 and is comparable to Benchmark 2 across tasks, and the x0-versus-noise comparison in Section 6.1 is instructive. However, those synthetic results do not establish the field-data claim, which is the distinctive contribution of the paper. The circularity in Algorithm 1 and Section 5.3 is concrete and easy to state precisely: the loss is computed against x_pseudo, and x_pseudo is generated by the model itself; the MSE in Figure 11 is measured against the model's own denoised output. The field benchmark comparisons are also not controlled, since Benchmark 1 and Benchmark 2 are not fine-tuned on field data while GSFM is. These issues do not require rejecting the paper; they require additional experiments or explicit caveats. The proposed concrete test—simulated field-shift with known targets, or independent external labels on real data—would settle whether the pseudo-label loop actually drives the model toward correct outputs or merely toward self-consistency. Given the reader's verdict is already CONDITIONAL and this concern directly supports conditionality, no change to the verdict is needed.","tokens_in":26102,"tokens_out":2493,"duration_ms":31509,"concrete_test":"Run Algorithm 1 on a simulated field-shift dataset: pretrain on the exact synthetic distribution from Section 4, then treat held-out elastic shot gathers generated from velocity models excluded from pretraining (with known clean, complete, broadband targets) as 'field' data. Track two MSE curves across fine-tuning stages 0, 1, 5, and 10: (a) MSE between model output and its own x_pseudo, and (b) MSE between model output and the known true target. If the pseudo-label MSE decreases while the true-target MSE does not improve—or worsens—the reported field-data gains reflect convergence to the model's own outputs rather than to the actual desired signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central field-data claim is that iterative fine-tuning (Algorithm 1) adapts GSFM to real seismic data and 'delivers significantly enhanced performance across diverse tasks.' The evidence for this claim is undermined by a circular pseudo-label loop. In Algorithm 1, x_pseudo is initialized in step 3 by the pre-trained model's own predictions on field data, and step 12 updates x_pseudo with the fine-tuned model's predictions. The loss in step 9 minimizes the distance to this self-generated target. Thus, improvements in the training objective only show that the model is converging to its own outputs. Section 5.3's quantitative evaluation is similarly circular: Figure 11 reports interpolation MSE against 'labeled data' defined as the denoised product of the very same GSFM (Figure 6b). A decreasing MSE across stages therefore does not establish that the model is recovering true missing signal; it may simply indicate increasing consistency with the model's own denoised estimate. Additionally, the field-data comparisons in Figures 6, 12, and 13 use Benchmark 1 and Benchmark 2 without any field-data fine-tuning, while GSFM is iteratively adapted on the exact field test shot gathers. This is an unfair comparison that cannot isolate the benefit of the pseudo-label strategy. Because the field evaluations are the primary support for the claimed generalization advantage, this circularity is load-bearing: without independent field labels or a controlled transfer experiment, the abstract's claim of 'significantly enhanced performance' on field data is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GSFM, a class-conditioned generative diffusion model intended to serve as an all-in-one seismic processing framework for denoising, backscattered noise attenuation, interpolation, and low-frequency extrapolation. The model is pre-trained on synthetic elastic shot gathers using an x0-prediction diffusion objective, then adapted to unlabeled field data through an iterative pseudo-label fine-tuning procedure (Algorithm 1). Synthetic experiments compare GSFM against a same-architecture supervised baseline (Benchmark 1) and a conventional pre-training-plus-fine-tuning baseline (Benchmark 2). Field experiments apply GSFM to a marine streamer dataset and compare against the two benchmarks, which are not adapted to field data. The paper also demonstrates uncertainty quantification via the standard deviation across stochastic samples.","tokens_in":26351,"tokens_out":4573,"duration_ms":46300,"significance":"If the field-data claims were validated, a single class-conditioned diffusion model that handles several seismic processing tasks and adapts to unlabeled field data without labeled ground truth would be a practical contribution, potentially reducing the need for task-specific models and providing useful uncertainty estimates. The strengths of the paper include the x0-targeted diffusion formulation, which enables single-step inference; the class-conditional multi-task architecture; the use of elastic synthetic modeling for pre-training; and the independent synthetic experiments, which reliably support the weaker claim that GSFM consistently outperforms Benchmark 1 and is competitive with Benchmark 2 on synthetic test data. The synthetic portion is generally well designed and reported. However, the field-data generalization claims are currently supported by evaluations that rely on the model's own outputs as pseudo-labels and as 'labeled data' in the interpolation MSE analysis, and the field comparison against benchmarks that are not fine-tuned on field data does not isolate the benefit of the proposed pseudo-label strategy.","major_comments":[{"comment":"Algorithm 1's iterative fine-tuning is circular with respect to correctness. Step 3 initializes x_pseudo with the pre-trained model's own predictions on field data, step 9 computes the loss against x_pseudo, and step 12 updates x_pseudo with the fine-tuned model's predictions. A decreasing training loss therefore only shows that the model is converging to its own outputs; it does not establish that the model is recovering the true signal on field data. The paper needs an external validation mechanism, such as independent field labels (e.g., manually quality-controlled outputs or paired well-log-constrained data), or a controlled experiment on held-out synthetic data that emulates the field distribution shift, to distinguish self-consistency from accuracy.","section":"Algorithm 1, Section 5.2"},{"comment":"The quantitative interpolation evaluation on field data uses the fine-tuned GSFM's own denoised output (Figure 6b) as the 'labeled data' against which MSE is computed. Figure 11 reports MSE trends across fine-tuning stages for this self-generated reference, and Figures 14 and 15 compare prediction means to the same denoised product. Consequently, the decreasing MSE across stages can be explained by the model becoming increasingly consistent with its own denoised estimate, rather than by more accurate interpolation of the missing signal. An independent reference for the missing traces (e.g., reserved traces from complete recordings, or a separate denoising method applied to the same field data) is needed to support the claim of improved interpolation accuracy.","section":"Section 5.3, Figure 11, Figure 14, Figure 15"},{"comment":"The field-data comparison is unfair because GSFM is iteratively fine-tuned on the exact field test shot gathers while Benchmark 1 and Benchmark 2 are used as pre-trained synthetic-only models without any field-data adaptation. The observed improvement could be attributable to any adaptation mechanism, not specifically to the pseudo-label strategy. A controlled comparison in which a benchmark is fine-tuned on the same field data with the same iteration budget (or, alternatively, GSFM is evaluated without field fine-tuning) is necessary to claim that the iterative pseudo-label strategy is the source of the field-data gains.","section":"Section 5.1, Figures 6, 12, 13"},{"comment":"The abstract claims that GSFM 'surpasses benchmarks with equivalent architectures in all tasks,' but the reported synthetic results do not support this. Table 2 shows Benchmark 2 achieves lower MSE than GSFM at 20%, 30%, 40%, and 50% noise levels, and Table 3 shows Benchmark 2 lower at 20% through 50% missing levels. Section 4.5 reports that Benchmark 2 outperforms GSFM on low-frequency extrapolation (MSE 3.11e-07 vs. 6.0e-07), and Section 4.6 acknowledges this. The claim should be revised to state that GSFM consistently outperforms Benchmark 1 and is competitive with, or slightly behind, Benchmark 2 on some synthetic tasks, reserving any 'surpasses' claim for the specific head-to-head settings where the data support it.","section":"Abstract, Section 4.2-4.5, Tables 2 and 3"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'DSFM' in figure captions (Figures 2-5, 7-13), 'steamer survey' for 'streamer survey' in Section 5.1, 'finnaly' in Section 5.4, 'gudie' in Section 7, 'weak' for 'weak' in Section 1, and 'he' for 'the' in the caption of Figure 6. The manuscript would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The text references 'Residual blocks (Figure 1d)' and 'Attention block (Figure 1e)', but the caption of Figure 1 lists '(e) The residual block. (d) The attention block.' The panel labels in the figure or the text references should be corrected for consistency.","section":"Section 3.5, Figure 1"},{"comment":"The MSE values for the noise-targeted GSFM are non-monotonic in the number of sampling steps, with some entries dramatically worse at T=1000 than at T=500 (e.g., Table 4, 10% noise: 2.50e-01 at T=1000 vs. 1.01e-05 at T=500; Table 5, 20% missing: 2.50e-01 vs. 1.27e-05). This is suggestive of training instability or an evaluation artifact; the authors should clarify whether the noise-targeted model was fully converged and whether these results are reproducible.","section":"Tables 4 and 5"},{"comment":"Table 1 appears to have swapped values in the rows for dt and nt: it lists 'dt Recording sampling step 376' and 'nt Number of recording timesteps 1.6e-2 s', which are dimensionally inconsistent with the stated total time of 6.016 s. The table should be corrected.","section":"Section 5.1, Table 1"},{"comment":"The term 'SSL' (self-supervised learning) is used for the fine-tuning strategy, but the training target is the model's own prediction on the same input, which is closer to self-training or pseudo-labeling than to classic self-supervised learning. The terminology should be clarified to avoid confusion with standard SSL approaches.","section":"Section 3.3, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central field-data claim rests on a self-distillation loop without external validation. The synthetic experiments are technically sound but support only a weaker claim than the abstract states. The field-data evaluation could be made rigorous by using an independent reference for the interpolation task (e.g., leaving out complete traces from a different acquisition pass or using a manual denoising reference), and by including a field-fine-tuned benchmark as a control. These changes are within the scope of a major revision and would substantially increase the credibility of the generalization claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is that a single class-conditioned diffusion model can handle denoising, backscattered-noise attenuation, interpolation, and low-frequency extrapolation with a shared architecture, and the synthetic results back that up. GSFM beats the plain U-Net baseline consistently and is competitive with the fine-tuned per-task baseline, sometimes better at high degradation levels. The x0-target prediction with single-step sampling is a practical efficiency win, and the uncertainty-quantification demo is a nice touch. That is a solid contribution to exploration seismology's deep-learning toolbox.\n\nThe soft spots are real but not fatal. First, the abstract claims GSFM 'surpasses benchmarks with equivalent architectures in all tasks,' but Tables 2 and 3 show Benchmark 2 winning at intermediate noise and missing levels. The body is more honest about this, so the abstract should be fixed. Second, the field-data evaluation is circular in a way that matters. Algorithm 1 fine-tunes against pseudo-labels produced by the model itself, and Figure 11 evaluates interpolation against the model's own denoised output as 'labeled data.' A dropping MSE across stages mostly confirms convergence to the model's own estimate, not recovery of true signal. Third, the field-data comparison pits GSFM, which was iteratively adapted on those exact shot gathers, against benchmarks that were never fine-tuned on field data at all. That comparison cannot isolate the benefit of the pseudo-label strategy. Code is also private during review, so the synthetic results are not independently checkable right now.\n\nNone of this sinks the paper, because the synthetic experiments stand on their own and the multi-task diffusion formulation is new. But the field-data claims need fair baselines (e.g., benchmarks fine-tuned with the same iteration budget on field data), and ideally some external validation beyond self-generated labels. The authors also need to release the code.\n\nThis deserves a serious referee. I would send it out with a request for major revision focusing on the abstract, the field-data evaluation, and code availability. A careful reader can learn from the synthetic results and the framework design even if the field-data section stays where it is.","headline":"A genuinely useful multi-task diffusion framework with solid synthetic experiments, but the field-data claims are built on a pseudo-label loop and an overreaching abstract; referee with a request for fair baselines.","tokens_in":26913,"tokens_out":1631,"would_cite":true,"duration_ms":18678,"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":"One generative diffusion model, GSFM, can simultaneously denoise, interpolate, and widen the bandwidth of seismic data, and the authors show that iterative self-supervised fine-tuning lets it transfer from synthetic training to unlabeled…","keywords":["generative diffusion model","seismic processing","foundation model","multi-task learning","interpolation","low-frequency extrapolation","self-supervised fine-tuning","uncertainty quantification"],"falsifier":"Start from a field dataset with independent ground truth, e.g., data whose true clean component is known from a reliable non-learning method or a controlled acquisition. Run the iterative fine-tuning and, at each stage, compute the mean squared error of the model output against both the independent truth and the model's own pseudo-labels. The central claim fails if the error to the independent truth stops decreasing while the error to the pseudo-labels keeps dropping.","tokens_in":25854,"feed_emoji":"🌊","tokens_out":5998,"duration_ms":55434,"temperature":0.7,"pith_summary":"The paper proposes GSFM, a single generative diffusion model that performs four seismic processing tasks at once: denoising, backscattered noise attenuation, interpolation, and low-frequency extrapolation. The central claim is that one model, pre-trained on synthetic shot gathers and fine-tuned on unlabeled field data, can match or beat task-specific networks that are trained or fine-tuned separately. If true, seismic processing could move from a chain of independently tuned algorithms to one shared model, and the same model's sampling randomness would provide a per-pixel uncertainty map of its own output. The authors test this against two same-architecture benchmarks on synthetic data and on a marine field dataset, where their iterative fine-tuning strategy outperforms the alternatives.","feed_headline":"One diffusion model cleans, fills, and widens seismic data","feed_subtitle":"The same model denoises, attenuates backscatter, interpolates, and recovers low frequencies on synthetic and field data.","key_machinery":"The central object is a dual-channel denoising diffusion model: channel one holds a noisy version of the target shot gather (labeled $x_0$), channel two holds the degraded data specific to the task, and a task label $c \\in \\{0,1,2,3\\}$ is embedded alongside the time step. The network is trained to predict $x_0$ directly with the loss $\\mathbb{E}\\|x_0 - x_{0,\\theta}(x_t, x, t, c)\\|^2$, which lets inference use a single DDIM reverse step. The fine-tuning machinery is Algorithm 1, an iterative pseudo-labeling loop: the pre-trained model predicts field data, those predictions are corrupted by the task-specific degradation to form the conditioning channel, and the model is fine-tuned against them, with pseudo-labels regenerated at each of $S$ stages.","core_discovery":"The authors claim that a generative diffusion model trained to predict the clean, complete, broadband target directly, rather than the added noise, can act as a multi-task seismic processing engine. The model takes two input channels: a noisy version of the ideal target and the task-degraded data, plus a class label that tells it which task is being asked. After pre-training on synthetic data, an iterative self-supervised loop refines the model on field data: the model's own predictions become pseudo-labels, are corrupted with the task-specific degradation, and the model is fine-tuned against them, with the pseudo-labels refreshed at fixed intervals. In their experiments the pre-trained model matches a conventionally fine-tuned benchmark on synthetic data, and the iterative strategy beats both benchmarks on field backscattered-noise attenuation, interpolation, and low-frequency extrapolation.","pith_inferences":["If the pseudo-label loop is trustworthy, the same synthetic-pre-training-plus-iterative-self-distillation recipe should transfer to other label-scarce geophysical tasks such as multiple removal and ground-roll suppression.","The uncertainty maps could be used as acquisition quality control: high-uncertainty patches would flag where additional receivers or source effort is needed, independent of any downstream imaging step.","A sharper test of the pseudo-label assumption would be to fine-tune with a small set of manually verified labels and check whether the model's uncertainty only drops in places where those labels agree with its predictions."],"forward_implications":["One GSFM instance can replace four separately trained networks, cutting both training time and the need for task-specific labeled data.","Because sampling starts from random noise, repeated predictions give a standard-deviation field that marks unreliable output regions; the paper shows this uncertainty correlates with residual error in interpolation.","The iterative pseudo-label fine-tuning is a recipe for adapting any synthetic-pre-trained diffusion model to unlabeled field data, not just for the four tasks demonstrated.","Direct $x_0$ prediction keeps accuracy at a single sampling step, so the generative model's inference cost is comparable to a conventional feed-forward network."],"supporting_citations":[{"why":"Supplies the DDPM forward process and reparameterization that GSFM's dual-channel noisy input is built on.","marker":"[Ho et al., 2020]"},{"why":"Supplies the DDIM deterministic reverse sampling that GSFM uses for efficient one-step inference.","marker":"[Song et al., 2020]"},{"why":"Provides the x0-prediction objective that lets GSFM train on target prediction and sample in a single step.","marker":"[Bansal et al., 2024]"},{"why":"The prior application of GDMs to seismic processing that GSFM extends from synthetic to field data and from single-task to multi-task.","marker":"[Durall et al., 2023]"},{"why":"The pre-training and fine-tuning paradigm for seismic processing that GSFM builds on and compares against with the same architecture.","marker":"[Harsuko and Alkhalifah, 2022]"},{"why":"The Seismic Foundation Model benchmark, a masked-autoencoder foundation model that GSFM compares against as a traditional pre-training strategy.","marker":"[Sheng et al., 2023]"},{"why":"Supplies the workflow for generating random elastic velocity models used to create the synthetic pre-training dataset.","marker":"[Ovcharenko et al., 2022]"},{"why":"Supplies the Deepwave elastic modeling engine used to simulate the synthetic shot gathers for pre-training.","marker":"[Richardson, 2023]"}],"fun_headline_variants":["One diffusion model handles four seismic tasks in a single pass","Unified generative diffusion model for all seismic processing tasks","Target-oriented diffusion: one model, four seismic tasks","Generative foundation model solves four seismic processing tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fine-tuning loop treats the model's own field-data predictions as trustworthy training targets; if those pseudo-labels carry systematic errors, the fine-tuned model will learn to reproduce them instead of the true signal.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion model handles four seismic tasks in a single pass","Unified generative diffusion model for all seismic processing tasks","Target-oriented diffusion: one model, four seismic tasks","Generative foundation model solves four seismic processing tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001193,"raw_usage":{"total_tokens":4921,"prompt_tokens":943,"completion_tokens":3978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3915}},"tokens_in":559,"tokens_out":3978,"duration_ms":28132,"temperature":1.0,"reasoning_tokens":3915,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:32:23.008944+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Start from a field dataset with independent ground truth, e.g., data whose true clean component is known from a reliable non-learning method or a controlled acquisition. Run the iterative fine-tuning and, at each stage, compute the mean squared error of the model output against both the independent truth and the model's own pseudo-labels. The central claim fails if the error to the independent truth stops decreasing while the error to the pseudo-labels keeps dropping.","supporting_citations":[{"cited_title":"Deep diffusion models for seismic processing","cited_arxiv_id":null,"evidence_quote":"The prior application of GDMs to seismic processing that GSFM extends from synthetic to field data and from single-task to multi-task."},{"cited_title":"Storseismic: A new paradigm in deep learning for seismic processing","cited_arxiv_id":null,"evidence_quote":"The pre-training and fine-tuning paradigm for seismic processing that GSFM builds on and compares against with the same architecture."}],"review_version":1}