REVIEW 3 major objections 6 minor 3 references
Physics-Driven Self-Supervised Deep Learning for Free-Surface Multiple Elimination
T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read A neural network can learn to remove free-surface multiples from seismic data without labeled examples, using only a physics-based reconstruction loss.
desk verdict Novel self-supervised SRME formulation with a genuine trivial-solution flaw; the idea deserves a serious referee but the current evidence does not support the central claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the factorized wavefield relation $P = P_0(I + AP)$ (Equation 2), restated from SRME, together with the reconstruction loss (Equation 3). Here $P$ is the measured full wavefield, $P_0$ is the primary-only wavefield, and $A$ is the surface operator $A = I R S^{-1}$, combining surface reflectivity $R$ and source wavelet $S$. The network maps $P$ to $(\hat{P}_0, \hat{a})$; plugging those into the relation reconstructs $\hat{P}$, and the loss compares it to $P$. This turns multiple elimination into a self-supervised factorization problem: no labeled primaries are needed, because the physical equation acts as the training supervisor.
What would settle it
On synthetic data with known true primaries, train the network from several random initializations and compare each estimated primary wavefield with the true one; if any low-loss solution yields $\hat{P}_0$ far from the true primaries or $\hat{A} \approx 0$, then the reconstruction loss alone does not identify the primaries.
Extended reading notes
Core claim
The paper's central claim is that the free-surface multiple-free wavefield can be parameterized directly by a deep network trained only on the full wavefield, provided the loss function encodes the physics of surface-related multiple generation. In this framework, one encoder and two decoders estimate the primaries $\hat{P}_0$ and a one-dimensional, spatially invariant surface operator $\hat{a}$. These outputs are fed through Equation (2), $P = P_0(I + AP)$, to form a reconstructed full wavefield, and the MSE between reconstruction and input is the entire training signal. The authors argue that the governing equation alone supplies enough constraint for the network to learn a useful separation, and they demonstrate on synthetic and field examples that the resulting primaries show better multiple suppression and primary preservation than the two SRME-based benchmarks, while acknowledging a higher computational burden.
Load-bearing premise
Everything rests on the assumption that minimizing the reconstruction error in Equation (3) leads the network to a physically meaningful factorization rather than the trivial one where the estimated primaries equal the input and the surface operator is zero; the paper does not constrain, regularize, or analyze this degeneracy.
Editorial extensions
If this is right
- Free-surface multiple elimination becomes a label-free inverse problem, and the same physics-loss recipe can be applied to other wavefield-separation tasks with known governing equations.
- The SRME benchmarks require adaptive subtraction with parameter tuning, while the proposed network avoids that tuning step at the price of per-dataset retraining on a GPU.
- If the factorization is reliable, the estimated surface operator $\hat{a}$ may itself be a usable byproduct, encoding information about surface reflectivity and the source wavelet.
- Extending from one dataset per training run to multi-dataset or cross-gather training is the stated next step; success there would make the method practical for large surveys.
Reading between the lines
- The minimization of Equation (3) admits a trivial exact solution $\hat{P}_0 = P$ and $\hat{A} = 0$; the paper does not analyze why training avoids it, so reproducibility across initializations and architectures is a critical untested condition.
- A quantitative comparison on synthetic data, where true primaries are known, would sharpen the claim of outperforming SRME, which the paper supports mainly through visual inspection of seismic sections.
- Because the reconstruction loss checks consistency rather than physical uniqueness, the network could converge to different factorizations that all fit the data; adding sparsity or a multiple-prediction term could stabilize the solution.
- The same self-supervised structure could be adapted to 3D acquisition and to tasks like deghosting or deblending, but scaling and compute remain open.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a physics-driven self-supervised deep-learning method for free-surface multiple elimination. A 3D CNN is trained to output both an estimated primary wavefield and a surface operator, using only a reconstruction loss based on the SRME forward equation P = P0(I + AP). The authors claim the method outperforms Global LS-SRME and Local LS-SRME on one synthetic and one field dataset, based on visual comparison. The central claim is that the network learns a physically meaningful primary estimate without any labeled data.
Significance. If the method is shown to be well-posed and validated quantitatively, the idea of embedding the SRME physics directly into a self-supervised loss is attractive and could reduce reliance on labeled or synthetic training data. The paper demonstrates a potentially useful application of physics-driven deep learning to seismic processing. However, the current manuscript does not establish that the learned factorization is physically meaningful: the loss function is degenerate, the evaluation is purely visual, and no quantitative evidence is provided. Thus the significance remains conditional on substantive revisions.
major comments (3)
- [The Proposed Physics-Driven Learning Method (Eq. (3))] The loss in Eq. (3) admits trivial zero-error solutions. Setting the estimated surface operator a_hat = 0 and the estimated primary p0_hat = p yields p_hat = p and J = 0. More generally, for any A for which (I + AP) is invertible, choosing P0 = P(I + AP)^{-1} exactly reconstructs P. The manuscript provides no constraint, regularization, or analysis of initialization or training dynamics to exclude these non-physical branches. Since the central claim is that the network estimates primaries, the authors must either add a constraint that breaks this degeneracy or empirically demonstrate that the trained network avoids it, for example by reporting the estimated A and showing that the estimated primaries differ from the input in the expected way. As written, Eq. (3) enforces only consistency with Eq. (2), which is an identity for infinitely many (P0, A) pairs, so the loss alone does not identify a physically meaningful primary.
- [Synthetic Data Examples and Field Data Example] The claim that the proposed method outperforms Global LS-SRME and Local LS-SRME is based solely on visual comparison of one synthetic and one field example. No quantitative metrics are reported: for the synthetic data, where true primaries are known, one could compute SNR or residual multiple energy; for the field data, one could compare the predicted multiples with a reference prediction or measure energy in a multiple-only window. Quantitative evaluation is especially important given the trivial-solution issue above: a model that simply outputs the input as primaries would show no 'primary leakage' in a visual comparison while failing to remove multiples. The authors should provide numeric measures of estimation accuracy and, for the field data, a demonstration that the learned surface operator is non-trivial and that the eliminated multiples have plausible physical characteristics.
- [Proposed Method / Experiments] The manuscript does not report the estimated surface operator a_hat or its properties (e.g., its length, amplitude spectrum, or spatial variation). The surface operator is a key physical output of the network; without showing it, the reader cannot assess whether the learned A corresponds to a physically reasonable free-surface operator or is an arbitrary filter that happens to satisfy the reconstruction. Reporting the learned A, at least for the synthetic case where the true source wavelet and reflectivity are known, would substantially strengthen the claim that the network has learned physics rather than an autoencoding factorization.
minor comments (6)
- [SRME Theory] There is a typo: 'framewrok' should be 'framework'.
- [Synthetic Data Examples, Figure 2 caption] The caption states 'estimated primaries from Global LS-SRME, Local LS-SRME, and the proposed method, respectively' but then refers to '(c) to (f)'; only (c), (d), and (e) are present. Please correct the caption.
- [Eq. (3)] The summation limits in Eq. (3) are typeset in a garbled way and should be written using standard notation, e.g., J = (1/(n_s n_r n_t)) sum_{i,j,k} |p(i,j,k) - p_hat(i,j,k)|^2.
- [Proposed Method] Please provide the training hyperparameters (optimizer, learning rate, batch size, initialization scheme, number of layers, filter sizes, and data normalization) to enable reproducibility. The current description gives only the filter-count progression.
- [Field Data Example] It would be helpful to specify how the field data were preprocessed (e.g., deghosting, designature, regularization) and how the SRME benchmarks were configured (window lengths, filter lengths) so that the comparison is reproducible.
- [Discussions] There is a typo: 'gnerateing' should be 'generating'.
Circularity Check
The self-supervised loss admits the trivial factorization A=0, P0=P with zero error, so the 'predicted' primaries are fitted to reconstruct the input through Eq. (2) rather than constrained by independent physics.
-
self definitional
[The Proposed Physics-Driven Learning Method, Eqs. (2) and (3), Figure 1]
"Instead, they are passed into the governing equation linking the full wavefield, primaries, and multiples ... This physical equation specifically refers to equation (2) ... Its output ... is expected to be a reconstruction of the full wavefield data p̂. ... the specific loss function employed in this work is the mean square error (MSE), ... J = ... ΣΣΣ (p(i,j,k)−p̂(i,j,k))^2"
The loss J is the MSE between the input p and the reconstruction p̂ built from the two network outputs p̂_primary and â via Eq. (2). Since both outputs are free variables, the reconstruction condition is an identity that can always be solved: setting â=0 and p̂_primary=p gives p̂=p and J=0; more generally, for any â with (I+âp) invertible, p̂_primary=p(I+âp)^{-1} also gives J=0. Thus zero reconstruction error does not single out free-surface multiples; it only requires a factorization of the input. The model's 'estimated primaries' are therefore fitted values that reproduce the input through the same equation (2) that defines the target, not predictions constrained by independent physics.
full rationale
The paper's derivation chain is: Eq. (2), P=P0(I+AP), is called the 'governing physical equation'; the CNN outputs P0 and A; these are fed back into Eq. (2) to reconstruct P; the loss (3) measures reconstruction error. Because P0 and A are both unconstrained outputs, Eq. (2) is not a physical constraint that determines P0; it is a tautology that always has a solution for any P and any invertible (I+AP). In particular, A=0, P0=P gives zero loss. Therefore the network is not forced to learn anything about free-surface multiples; the success of primary estimation rests on an unexamined inductive bias of the 3D CNN and its training procedure. The qualitative comparisons with SRME benchmarks are external and do not break this circularity: they are not part of the training signal and do not demonstrate that the loss itself selects primaries. No load-bearing self-citation is present: Verschuur et al. (1992) is the standard SRME foundation and Wang et al. (2017) is only cited as an example of a derivative method. Thus the circularity is in the construction of the loss/target relation, not in citation practice. Score 6: the central claim of 'physics-driven' primary prediction reduces to fitting a factorization of the input under the defining equation.
Assumptions & free parameters
free parameters (4)
- 3D CNN encoder-decoder weights =
not reported (optimized per dataset)
- Surface operator filter a_hat =
not reported (estimated per dataset)
- Training hyperparameters (optimizer, learning rate, batch size, initialization) =
not reported
- Network architecture choices (filter counts 64-1024, strides, fully connected layers) =
chosen by authors
assumptions (4)
- domain assumption The SRME forward model P = P0(I + A P) is an exact description of the full wavefield for densely sampled 2D data.
- domain assumption The input data satisfy fully dense sampling with equal numbers of sources and receivers.
- domain assumption A single 1D, spatially invariant surface operator filter is sufficient to represent the free-surface effect for each dataset.
- ad hoc to paper The CNN's architecture and initialization bias the optimization away from the trivial solution P0=P, A=0 and toward physically meaningful primaries.
Cite this review
Pith. "Pith review of Physics-Driven Self-Supervised Deep Learning for Free-Surface Multiple Elimination." pith.science (2026). https://pith.science/paper/JU3MV6JY
@misc{pith2026250205189,
author = {Pith},
title = {Pith review of: Physics-Driven Self-Supervised Deep Learning for Free-Surface Multiple Elimination},
year = {2026},
howpublished = {\url{https://pith.science/paper/JU3MV6JY}},
note = {Machine review of arXiv:2502.05189}
}
read the original abstract
In recent years, deep learning (DL) has emerged as a promising alternative approach for various seismic processing tasks, including primary estimation (or multiple elimination), a crucial step for accurate subsurface imaging. In geophysics, DL methods are commonly based on supervised learning from large amounts of high-quality labelled data. Instead of relying on traditional supervised learning, in the context of free-surface multiple elimination, we propose a method in which the DL model learns to effectively parameterize the free-surface multiple-free wavefield from the full wavefield by incorporating the underlying physics into the loss computation. This, in turn, yields high-quality estimates without ever being shown any ground truth data. Currently, the network reparameterization is performed independently for each dataset. We demonstrate its effectiveness through tests on both synthetic and field data. We employ industry-standard Surface-Related Multiple Elimination (SRME) using, respectively, global least-squares adaptive subtraction and local least-squares adaptive subtraction as benchmarks. The comparison shows that the proposed method outperforms the benchmarks in estimation accuracy, achieving the most complete primary estimation and the least multiple energy leakage, but at the cost of a higher computational burden.
Reference graph
Works this paper leans on
-
[1]
The framework of the physics-driven self-supervised learning strategy with a visualization of the employed 3D CNN model. The governing physical equation,which links the full wavefield, primaries and multiples, serves as the foundation of the geophysical primary estimation method SRME. The input data p is a 3D cube representing the full wavefield, with dim...
work page 2016
-
[3]
Faculty Of Civil Engineering and Geosciences, Delft University of Technology, The Netherlands Summary In recent years, deep learning (DL) has emerged as a promising alternative approach for various seismic processing tasks, including primary estimation (or multiple elimination), a crucial step for accurate subsurface imaging. In geophysics, DL methods are...
work page 2016
-
[1992]
From a geophysical perspective, multiples can be predicted by the convolution of the full wavefield. The established practice of SRME methods consists of two parts: first, predicting the multiples, and second, subtracting the predicted multiples from the full wavefield. Effectiveness depends heavily on accurately matching the predicted multiples to the or...
work page 1992
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.