{"id":"35831a35-17f2-4814-97f8-6a15fdce4f1a","arxiv_id":"2501.15473","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A semi-supervised AVO inversion method adds a masked-trace reconstruction constraint so the network learns lateral structure from sparse well-log locations.","lead":"The paper combines three neural networks to estimate rock properties from seismic data, using a new training rule that forces neighboring traces to help reconstruct well-log measurements. If the improvement holds up, it would make seismic reservoir imaging sharper with fewer drill holes for calibration.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may come from the low-frequency prior rather than the SSFC mechanism: baselines' access to the low-frequency model is unstated, and the field blind-well description is internally inconsistent.","rationale":"The reader identified the low-frequency prior as the weakest assumption; I agree that this is the right area but sharpen it: the load-bearing failure is not primarily that the prior's error would dominate away from wells (a real but standard limitation of all such inversions), but that the experimental comparison does not control the prior across methods. Eq. 17 feeds mlow to the inversion network at every trace, and mlow is highly informative: in the synthetic test it is the true model's low-frequency component, and in the field test it is produced by interpolating and smoothing the actual well logs on that line. If the 1D-SL/1D-SSL/2D-SSL baselines were trained without this input — which the paper never states — the accuracy gap in Tables II and III is expected even without any benefit from the SSFC mechanism, because band-limited AVO data do not constrain the low-frequency trend of VP, VS, and density. The missing ablation (no run of SSFC-SSL without mlow, and no run of the baselines with mlow) makes the attribution of the improvement to the 'strong spatial feature constraints' unverifiable as written. The field protocol's internal inconsistency about the blind well (G3 in one sentence; G2 in Fig. 21 and Table III) corroborates that the prior/training split was not handled with care: if the evaluation well's logs contributed to the low-frequency model, the field-data demonstration is circular. I do not therefore move to REJECT: the framework is plausible, the losses are reported to converge, the synthetic experiment is internally coherent, and the remedy is a standard controlled experiment plus a clean protocol statement. Minor issues — Eq. 21's PCC denominator uses 'yk' where 'mk' is intended, the duplicate 'where' before Eq. 8, and the text citing SSIM where Table III reports RMSE — are cosmetic. The reader's CONDITIONAL verdict stands, with the added explicit condition that the controlled-prior comparison and blind-well clarification be provided; hence UNCHANGED with 'partial' agreement, since I weight the comparison confound more heavily than the prior-inaccuracy-away-from-wells concern emphasized in the reader's wording.","tokens_in":19633,"tokens_out":21247,"duration_ms":188049,"concrete_test":"Controlled-prior re-run: retrain 1D-SL, 1D-SSL, and 2D-SSL with the identical low-frequency model mlow as conditional input (same normalization, same 5-trace windows, same hyperparameters), and retrain SSFC-SSL without the mlow input while keeping all other losses fixed. If the SSFC-SSL advantage over 2D-SSL in Table II collapses to within seed-level variance (or reverses when mlow is removed), the headline improvement is attributable to the prior rather than to the strong spatial feature constraint. Include the same run in the field test; as part of the check, confirm which well is truly held out and exclude its logs from both training and the low-frequency-model construction, since the text alternately names G3 and G2 as the blind well.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: SSFC-SSL improves lateral continuity and accuracy over 1D/2D DL baselines. For this claim to be established, the comparison must control for the low-frequency prior that the proposed method injects. In Eq. 17 the inversion network takes (D, mlow) as input at every trace; on synthetic data mlow is the true model low-pass filtered (Fig. 7f-h), and on field data it is obtained by layer interpolation, extrapolation, and smoothing of the well-log curves (Section III-C). The paper never states whether the 1D-SL, 1D-SSL, and 2D-SSL baselines also received mlow. If they did not, the improvements in Tables II and III (e.g., P-wave RMSE 0.0420 vs 0.0553 for 2D-SSL) may be largely an unequal-information effect: the prior already supplies the low-frequency answer, which is the hardest part of band-limited AVO inversion to recover from seismic alone, and the SSFC network's specific contribution is not demonstrated. Two details aggravate the concern. First, the prior construction (low-pass cutoff, interpolation/smoothing parameters) is unreported, so the experiment cannot be reproduced as written. Second, the field-data protocol is internally inconsistent: the text says 'Wells G1, G2, and G3 were used for training, while well G3 was used to evaluate,' but Fig. 21 and Table III evaluate 'the blind well G2.' If the evaluation well's logs entered the training set or the prior, the field metrics are not a blind test. The concern is checkable and does not require assuming bad faith; a controlled comparison would settle whether the SSFC mechanism or the prior drives the reported gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SSFC-SSL, a semi-supervised AVO inversion method that combines an inversion network, a forward network, and a strong spatial feature constraint (SSFC) network. The method uses a label-annihilation operator to constrain predictions at well locations from adjacent non-well traces and introduces a self-consistency loss, while also feeding a low-frequency model of elastic parameters as conditional input at every trace. The authors compare SSFC-SSL with 1D-SL, 1D-SSL, 2D-SSL, and model-based (L-BFGS) methods on a Marmousi2 synthetic example and on field data from the South China Sea, reporting improved PCC, R², RMSE, and visually smoother lateral continuity.","tokens_in":19973,"tokens_out":5999,"duration_ms":48283,"significance":"If the empirical claims are supported, the SSFC mechanism would be a useful contribution to deep-learning-based AVO inversion, particularly under scarce well-log labels, and the paper provides a clear architecture, explicit loss functions, and both synthetic and field demonstrations. The proposed two-way spatial constraint idea is interesting and potentially valuable. However, the reported gains currently lack a controlled comparison: the low-frequency prior is not shown to be equally available to the baselines, there is no ablation isolating the SSFC loss, single-run metrics are reported without error bars or significance tests, and the field evaluation protocol contains an internal inconsistency. These issues must be resolved before the comparative claim can be accepted.","major_comments":[{"comment":"The manuscript never states whether the 1D-SL, 1D-SSL, and 2D-SSL baselines received the low-frequency model mlow as input during training and inference. Since the inversion network in Eq. (17) takes (Dl, mlow) and (Du, mlow) as inputs, the proposed method injects low-frequency information that is known to be the most difficult part of band-limited AVO inversion to recover from seismic alone. If the baselines operate on seismic data without mlow, the improvements in Tables II and III (e.g., P-wave RMSE 0.0420 vs. 0.0553 for 2D-SSL) may be an artifact of unequal information rather than the SSFC mechanism. Please either provide mlow to all baselines or add an ablation of SSFC-SSL without mlow, and report both setups.","section":"III-B/III-C, Eq. (17)"},{"comment":"The field-data evaluation is internally inconsistent: the text states that 'Wells G1, G2, and G3 were used for training, while well G3 was used to evaluate the inversion results,' but Fig. 21 and Table III compare the inversion results 'at the location of the blind well G2.' If G2's logs were used in training or in constructing the low-frequency model via well-log interpolation and smoothing, then the reported field metrics are not a blind test. Please clarify which well is the blind test and confirm that its logs were excluded from both training and prior construction.","section":"III-C"},{"comment":"The construction of the low-frequency model is not described reproducibly. For the synthetic case, the text states only that 'a low-pass filter was applied to the model data to obtain the low-frequency model,' without specifying the cutoff frequency, filter order, or design. For the field case, it states that the low-frequency information is obtained by 'layer interpolation, extrapolation, and smoothing of the corresponding well-log curves,' without giving the interpolation or smoothing parameters. These details are essential because mlow is a conditional input at every trace and can dominate the inversion result; without them, the experiments cannot be reproduced or compared fairly.","section":"III-B/III-C, Eq. (17)"},{"comment":"The central claim that SSFC-SSL 'significantly improves' inversion accuracy is supported only by single-run metrics. There are no repeated runs, error bars, or statistical significance tests, and no ablation that removes the SSFC loss (LSSFC and LSC) or the low-frequency model. Given that the improvement over 2D-SSL is modest for some parameters (e.g., density RMSE 0.0123 vs. 0.0166 on the synthetic data), run-to-run variation could change the conclusion. Please add repeated runs with error bars and an ablation study to isolate the contribution of each loss term and of mlow.","section":"III-B, Tables II-III"},{"comment":"The paper repeatedly claims improved 'lateral continuity,' but no metric for lateral continuity is reported; the tabulated metrics (PCC, R², RMSE) are pointwise measures. Please quantify lateral continuity directly, for example with SSIM, total variation, or a trace-to-trace smoothness measure, to substantiate the claim.","section":"III-B, Figs. 10-16 and Tables II-III"}],"minor_comments":[{"comment":"The text says Table III reports 'PCC, R2, and SSIM values,' but the table actually reports PCC, R², and RMSE; please correct the text or the table header.","section":"III-C"},{"comment":"The reference to 'Fig. 2(b)' in the discussion of 2D data-driven inversion should likely be 'Fig. 1(b)', since Fig. 2 shows the training process rather than the workflow schematic.","section":"I"},{"comment":"The subscript for the self-consistency loss is written as 'Lsc' with a lowercase 's' in Eq. (18), while Eqs. (15)-(17) use 'LSC'; please make the notation consistent.","section":"II-D, Eq. (18)"},{"comment":"The description of unlabeled data sampling as 'randomly sampled at an interval of 50 traces' is ambiguous: it could mean every 50th trace or a random selection with an average spacing of 50; please clarify.","section":"III-B"},{"comment":"The word 'significantly' in the abstract is not supported by statistical tests or error bars; please qualify the claim or add the corresponding statistical evidence.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a data/code availability statement, as the experiments are not reproducible as written due to the missing low-frequency-model construction details and the field-well inconsistency. The novelty relative to prior closed-loop and low-frequency-model-based inversion methods (e.g., Refs. [52]-[54]) should also be sharpened in the revision to clarify what is new beyond combining existing components."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. The core idea is a masked-reconstruction regularizer for 2D semi-supervised AVO inversion: after the inversion network predicts elastic parameters for a multi-trace patch, a label-annihilation operator zeroes the center trace, and an SSFC network tries to recover it from the neighboring traces; a self-consistency term stabilizes it. That combination is new in this literature, and it is a plausible way to propagate sparse well-log constraints laterally. The empirical work is also genuinely useful: the synthetic tests include very sparse label regimes (7 and 3 training traces out of 2720), and the proposed method beats the 1D/2D baselines on PCC/R2/RMSE consistently. The field data test is real, not a toy.\n\nThe soft spot is the comparison. Equation (17) feeds the inversion network the pair (D, mlow), where mlow is a low-frequency model. On synthetic data mlow is the true model low-pass filtered (Fig. 7f-h); on field data it is built from well-log interpolation, extrapolation, and smoothing. The paper never says whether the 1D-SL, 1D-SSL, and 2D-SSL baselines also receive mlow. If they do not, the reported gains in Tables II and III could come largely from the prior rather than from the SSFC mechanism — the low-frequency component is exactly the part of band-limited AVO inversion that is hardest to recover from seismic alone. This is a checkable issue, but it is load-bearing for the comparative claim. An ablation that removes mlow from SSFC-SSL, plus a baseline that also gets mlow, would settle it.\n\nThere is also an internal inconsistency in the field section. The text says wells G1, G2, G3 were used for training and well G3 was used for evaluation; the next paragraph and Fig. 21/Table III evaluate \"the blind well G2.\" Since G2 was listed as a training well, the field evaluation is not currently a clean blind test. That needs to be fixed before anyone relies on those numbers.\n\nSmaller complaints: no error bars or repeated runs, no ablation isolating the SSFC regularizer, and no released code or data. None of these are fatal on their own.\n\nNet: the paper is incremental but honest work, not a field redefinition. The method is sensible, the synthetic evidence is consistent, and the flaws are addressable in revision. I'd send it to peer review, but I'd insist on a controlled comparison of the low-frequency prior and a corrected explanation of which well was actually blind.","headline":"A credible but incremental masked-reconstruction regularizer for semi-supervised AVO inversion; the comparative claim is plausible, but the low-frequency prior is not controlled and the field blind-well description contradicts itself.","tokens_in":20478,"tokens_out":2824,"would_cite":false,"duration_ms":25843,"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":"The paper proposes a semi-supervised AVO inversion method whose label-annihilation and self-consistency losses create a two-way spatial mapping between well and non-well locations, and reports improved lateral continuity and accuracy over…","keywords":["AVO inversion","semi-supervised learning","strong spatial feature constraints","label annihilation operator","lateral continuity","low-frequency model","convolutional neural network","elastic parameter estimation"],"falsifier":"On a synthetic model with a strong lateral velocity contrast that the low-frequency interpolation intentionally smooths away, run SSFC-SSL and the 2D-SSL baseline without the low-frequency input; if SSFC-SSL's advantage over 2D-SSL shrinks below the paper's reported margin or disappears, the improvement is attributable to the low-frequency prior rather than the strong spatial feature constraint.","tokens_in":19450,"feed_emoji":"📈","tokens_out":6967,"duration_ms":57776,"temperature":0.7,"pith_summary":"The paper proposes a semi-supervised deep-learning method for prestack AVO inversion that recovers P-wave velocity, S-wave velocity, and density from multi-angle seismic gathers. Its central claim is that adding a strong spatial feature constraint between well and non-well locations, enforced by a label-annihilation operator and a recovery network, yields laterally continuous and more accurate parameter volumes than one- and two-dimensional deep learning baselines under scarce well-log labels. The method combines an inversion network, a forward-modeling network, and a strong spatial feature constraint (SSFC) network, and it feeds a low-frequency model of the elastic parameters as conditional input. On the synthetic Marmousi2 model and on a South China Sea field dataset, the paper reports the best Pearson correlation, R-squared, and RMSE among the compared methods, with the gains holding when only a fraction of traces carry labels.","feed_headline":"Two-way spatial constraints lift AVO inversion accuracy","feed_subtitle":"A label-annihilation operator makes traces away from wells learn from well data, tightening elastic-parameter images.","key_machinery":"The load-bearing object is the label-annihilation operator R, a matrix that keeps all traces except the well-location (middle) trace, setting that trace's entries to zero. The predicted parameter volume is multiplied by R, and the SSFC network is trained to reconstruct the zeroed trace from its neighbors under a self-consistency loss: the reconstruction must match the true well log where the label was annihilated (L_SSFC) and must reproduce the network's own prediction at the other traces (L_SC). This creates a two-way mapping: well labels constrain adjacent traces through the inversion network, and neighboring traces are used to regenerate the well trace, propagating lateral structure. The forward network provides a semi-supervised reconstruction loss on unlabeled seismic data, and a low-frequency elastic-parameter model is fed into the inversion network alongside the seismic gathers.","core_discovery":"The paper claims that AVO inversion can be improved by turning sparse well-log labels into a bidirectional spatial constraint. Instead of using well logs only to supervise the traces where wells exist, the method multiplies the predicted parameter volume by a label-annihilation matrix R that zeros out the well-location trace, then trains a second network to recreate that missing trace from its neighbors. Requiring the recreated trace to match both the well logs and the annihilated prediction makes spatial features flow from non-well to well locations and back, which is the strong two-way mapping that 1D and 2D baselines lack. In the reported experiments, this SSFC-SSL scheme achieves higher PCC and R-squared and lower RMSE for all three elastic parameters than the model-based L-BFGS, 1D supervised, 1D semi-supervised, and 2D semi-supervised comparisons, and it preserves accuracy when only 0.26% or 0.11% of traces carry labels.","pith_inferences":["An ablation that removes the low-frequency model input would isolate how much of the gain comes from the SSFC loss versus the prior; the paper does not report such an ablation.","The annihilation-plus-reconstruction recipe is generic: any sparse-label geophysical or image-reconstruction task could adopt it to propagate label information laterally, not just AVO inversion.","Because the SSFC network is trained on the inversion network's own predictions, an early inaccurate inversion output could teach the SSFC network a biased spatial prior; starting with a warm-up on well traces only might reduce that risk."],"forward_implications":["The method reduces the vertical striping artifacts that plague 1D trace-by-trace inversion, because the SSFC loss forces spatial coherence across adjacent traces.","Under scarce well-log labels, SSFC-SSL degrades more gracefully than 1D supervised learning; the paper reports it keeps the highest PCC and R-squared and lowest RMSE with 18, 7, or 3 training traces on the Marmousi2 model.","At prediction time only the inversion network is used; the SSFC and forward networks are discarded, so the improved lateral continuity is obtained without extra inference cost.","The authors state the framework is intended to extend to post-stack inversion, where the same spatial-constraint and semi-supervised losses would apply."],"supporting_citations":[{"why":"Supplies the 1D CNN baseline that lacks lateral continuity and defines the problem the paper attacks.","marker":"[37]"},{"why":"Establishes the 2D CNN approach that improves lateral stability, the direct baseline SSFC-SSL must beat.","marker":"[55]"},{"why":"Introduces semi-supervised RNN inversion via forward modeling, the strategy the paper combines with spatial constraints.","marker":"[43]"},{"why":"Provides the closed-loop semi-supervised inversion framework (forward network constrains inversion network) that SSFC-SSL extends.","marker":"[53]"},{"why":"Applies a closed-loop forward/inversion network specifically to AVO, the immediate predecessor of this method.","marker":"[54]"},{"why":"Introduces 2D semi-supervised inversion with domain adaptation, a key comparison for lateral continuity and accuracy.","marker":"[56]"}],"fun_headline_variants":["Well-label annihilation turns AVO inversion spatial","Bidirectional well-labels enhance AVO inversion","Semi-supervised AVO inversion gains lateral continuity","Two-way mapping from sparse well labels lifts AVO accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method requires a reliable low-frequency model of the elastic parameters, built by interpolating and smoothing well-log curves; if that model is wrong or missing far from wells, the inversion error will be dominated by the prior's error, and the spatial-constraint gains may vanish.","fun_headline_variants_meta":{"raw":{"variants":["Well-label annihilation turns AVO inversion spatial","Bidirectional well-labels enhance AVO inversion","Semi-supervised AVO inversion gains lateral continuity","Two-way mapping from sparse well labels lifts AVO accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000885,"raw_usage":{"total_tokens":3826,"prompt_tokens":955,"completion_tokens":2871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2810}},"tokens_in":571,"tokens_out":2871,"duration_ms":18858,"temperature":1.0,"reasoning_tokens":2810,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:14:59.412526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic model with a strong lateral velocity contrast that the low-frequency interpolation intentionally smooths away, run SSFC-SSL and the 2D-SSL baseline without the low-frequency input; if SSFC-SSL's advantage over 2D-SSL shrinks below the paper's reported margin or disappears, the improvement is attributable to the low-frequency prior rather than the strong spatial feature constraint.","supporting_citations":[{"cited_title":"Convolutional neural network for seismic impedance inversion,","cited_arxiv_id":null,"evidence_quote":"Supplies the 1D CNN baseline that lacks lateral continuity and defines the problem the paper attacks."},{"cited_title":"Deep learning for multidimensional seismic impedance inversion,","cited_arxiv_id":null,"evidence_quote":"Establishes the 2D CNN approach that improves lateral stability, the direct baseline SSFC-SSL must beat."},{"cited_title":"Semisupervised sequence modeling for elastic impedance inversion,","cited_arxiv_id":null,"evidence_quote":"Introduces semi-supervised RNN inversion via forward modeling, the strategy the paper combines with spatial constraints."},{"cited_title":"Seimic impedance inversion based on semi-supervised learning,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-loop semi-supervised inversion framework (forward network constrains inversion network) that SSFC-SSL extends."},{"cited_title":"Avo inversion based on closed-loop multitask conditional wasserstein generative adversarial net- work,","cited_arxiv_id":null,"evidence_quote":"Applies a closed-loop forward/inversion network specifically to AVO, the immediate predecessor of this method."},{"cited_title":"Seismic inversion based on 2d-cnns and domain adaption,","cited_arxiv_id":null,"evidence_quote":"Introduces 2D semi-supervised inversion with domain adaptation, a key comparison for lateral continuity and accuracy."}],"review_version":1}