Organizations: School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, Shaanxi, 710049, China. · Business School, Beijing Wuzi University, Beijing, 101149, China. · Geophysical Technology Research and Development Center, BGP, Zhuozhou, Hebei, 072751, China.
Seismic data interpolation must reconcile reconstruction fidelity with the computational budgets of acquisition quality control and subsequent imaging. Fixed-depth unfolding networks are optimized for a single operating point, whereas their intermediate estimates may be insufficiently trained for early deployment. We present MetaPOCS, a meta-learned multi-exit unfolding framework for budget-aware seismic interpolation. Building on a spatial--spectral POCS reconstruction backbone, MetaPOCS supervises intermediate stages and preserves the acquired traces at every exit. A Meta-GF-based optimization procedure constructs exit-oriented parameter displacements from a common initialization and learns their fusion weights through an equal-exit outer objective. The weights vary across parameter tensors and respect the nested support of the reconstruction stages. We derive the resulting fusion hypergradient to characterize how the outer objective reallocates competing updates, and distinguish this mechanism from scalar loss reweighting. At inference, one checkpoint provides a set of reconstruction budgets through a latency-calibrated exit policy. The evaluation framework couples matched-cost interpolation comparisons with finite-update diagnostics and a fixed reverse-time-migration operator, connecting multi-exit optimization to waveform and imaging fidelity. MetaPOCS thus provides a unified formulation for studying the reconstruction--computation trade-off of seismic unfolding across deployment budgets.
Figures & tables
Fig. 1: Detailed architecture of a single unrolled stage.
Fig. 2: MetaPOCS structure. Each stage combines an implicit time–space/Fourier prior with exact data consistency (DC), and each trained exit receives explicit time–space and spectral supervision. Reconstruction stages inherit SSI-POCS; multi-exit coordination and deployment are the focus here. The lower block operates during training. Inference runs a budget-selected prefix, without computing all exits and subsequently discarding their cost.
Fig. 3: MetaPOCS meta fusion.
Dataset
Shots
Receivers per Shot
Sample Time ( Nt )
Time Sample Interval ( Δt )
Spatial Sampling Interval ( Δs )
Train/val/test
SEGC3
45
201×201
625
8 ms
20 m
25000/10000/10000
MAVO
1001
120
1500
4 ms
25 m
10000/4000/6000
F3
600
440
951
4 ms
25 m
8000/3000/6000
Model94
277
480
2000
4 ms
15 m
10000/3000/3500
TABLE I: Datasets Information.
Control
Training / deployment
Isolated factor
SSI-POCS original
Original recipe, final exit
Relationship to the preceding reconstruction method
Fixed- d matched
Independent model at each d=2,4,6,8
Specialized training for each deployment budget
Deep-only truncated
Depth-8 supervision; prefix inference
Intermediate reconstruction without direct supervision
Equal multi-exit
Average identical exit losses
Direct supervision at several reconstruction depths
Scalar meta
Eq. ( 17 )
Global redistribution of exit loss weights
PCGrad / GradNorm
Common exits and objectives
Gradient-direction or gradient-scale balancing
TABLE II: Training and deployment controls. Matched comparisons share the reconstruction backbone, loss scaling, acquisition partitions, and masks.
Fig. 4: Panoramic view of physical iteration evolution across different unrolling mechanisms and supervision strategies (shot 474). The matrices visualize the step-wise wavefield reconstruction from initial iterations (T1, T3) to the final stage. Notably, traditional fixed-depth schemes (Rows 1-2) exhibit slow convergence, while the ‘Equal multi-exit‘ mechanism (Row 3) introduces severe gradient conflicts, blurring the intermediate wavefields (T3, T5). In contrast, the proposed ‘MetaPOCS‘ (Row 5), driven by dynamic bi-level meta-supervision, resolves early-stage feature entanglements rapidly, achieving unparalleled amplitude fidelity and structural continuity as early as T3, and yields the optimal final reconstruction.
Fig. 5: Quantitative stage-wise PSNR evolution trajectories under different supervision mechanisms (averaged across 8 test shots). The proposed MetaPOCS effectively overcomes the gradient conflicts observed in the Equal multi-exit baseline during early stages (T1–T4) and consistently maintains the highest reconstruction accuracy, demonstrating the superiority of the dynamic bi-level meta-supervision strategy.
Fig. 6: Qualitative comparison of reconstruction results on the SEG C3 dataset under 50% random trace removal (shot 12_4). We restore the seismic data to its original amplitude range.
Fig. 7: Qualitative comparison of F-K spectrum reconstruction results on the SEG C3 dataset under 50% random trace removal (shot 12_4). We restore the seismic data to its original amplitude range.
Fig. 8: Qualitative comparison of reconstruction results on the MAVO dataset under 50% trace removal. (a) Random undersampling (shot 180). (b) Consecutive large gaps (shot 697). We restore the seismic data to its original amplitude range.
Fig. 9: Qualitative comparison of F-K spectrum reconstruction results on the MAVO dataset under 50% trace removal. (a) Random undersampling (shot 180). (b) Consecutive large gaps (shot 697). We restore the seismic data to its original amplitude range.
Fig. 10: Qualitative comparison of reconstruction results on the MAVO dataset under 50% trace removal. (a) Random undersampling (shot 180). (b) Consecutive large gaps (shot 488). We restore the seismic data to its original amplitude range.
Fig. 11: Pareto efficiency analysis (PSNR vs. Parameters). The proposed method optimally occupies the top-left Pareto frontier, achieving the highest reconstruction fidelity with minimal parameter overhead.
Method
SEG C3 random
SEG C3 consecutive
MAVO random
MAVO consecutive
Curvelet
20.131/2.366e-02
1.0578/2.220e-01
24.600/1.292e-02
1.0598/2.333e-01
POCS-Net
22.486/1.664e-02
17.810 /3.174e-02
26.550/9.237e-03
23.178/1.660e-02
SP-Net
22.281/1.236e-02
15.578/3.875e-02
26.446/9.444e-03
19.110 /2.298e-02
D3UNet
24.304/1.366e-02
18.527/2.825e-02
27.840/8.138e-03
24.241/1.317e-02
MetaPOCS
28.302/8.600e-03
20.114/2.470e-02
32.977/5.657e-03
26.533/1.336e-02
TABLE III: Reconstruction comparison at 50% trace removal. Each entry is missing-region SNR (dB) / MAE; larger SNR and smaller MAE are preferable.
Method
Reconstruction Quality (dB)
Memory (MB)
Inf. Time (ms)
Q1
Q3
Q5
Q7
Allocated
Peak
Checkpoint
( T1 / T7 )
D3UNet
3.008
10.226
20.672
26.531
37.81
6630
9.27
3.783
Fixed- T
26.370
30.206
32.323
32.767
81.3
19124
5.90
1.567 / 8.033
Deep-only trunc.
14.275
19.420
23.066
32.202
81.3
19124
5.90
1.567 / 8.033
Equal-exit
24.937
30.765
31.630
31.919
72.60
17908
5.91
1.617 / 8.033
Exit-weight BLO
24.977
31.073
31.803
31.936
73.58
23676
5.91
1.617 / 8.033
TABLE IV: Reconstruction quality and computational cost across exit depths. Qd denotes missing-region SNR (dB). Memory and latency are measured on the same device and input geometry. Inference latency is reported as the minimum and maximum time corresponding to T=1 and T=7 , respectively. Training time includes meta optimization.
Prior
Loss recipe
MAE ↓
SSIM ↑
SNR ↑
Improvement
Spa–Spa, concat
L1
8.063e-03
0.941
28.587
✓
FFC–FFC, concat
L1
8.267e-03
0.943
28.588
✓
Spa–FFC, concat
L1
8.239e-03
0.946
29.549
✓
Spa–FFC, DFM
L1
6.862e-03
0.956
30.831
✓
Spa–FFC, DFM
L1+FFL
6.904e-03
0.957
31.049
✓
Spa–FFC, DFM
L1+sym+FFL
6.427e-03
0.959
31.463
✓
TABLE V: Interaction between implicit reconstruction priors, explicit reconstruction objectives, and training. Equal and meta columns report Q2/Q8 in dB; the final column gives their paired difference. Transform regularization is retained in every matched configuration.
Type
Model
MAE ↓
SNR ↑
SSIM ↑
PSNR ↑
SEGC3 to Model94: Random missing
D3UNet
0.0105
25.144
0.9501
31.712
Ours (T=7)
0.0081
26.788
0.9645
33.499
Ours (Best T=6)
0.0083
26.875
0.9645
33.586
MAVG to Model94: Random missing
D3UNet
0.0231
17.934
0.8136
24.502
Ours (T=7)
0.0149
21.057
0.9073
27.768
Ours (Best T=6)
0.0141
21.227
0.9073
27.938
TABLE VI: Generalization comparison of different models on the test set of the Model94 dataset (missing rate: 50%).
Deep unfolding turns iterative algorithms into trainable networks, combining model-based structure with learned components for efficient inference and optimization.
Large gaps in seismic acquisition disrupt reflection events and hinder spectral recovery. Existing seismic interpolation deep unfolding networks primarily rely on localized spatial priors, limiting their ability to aggregate the global context necessary for reconstructing these extensive missing regions. We propose a spatial-spectral interactive projection-onto-convex-sets network (SSI-POCS) to tackle this severely ill-posed problem. By unfolding the classical iterative solver, our network integrates physical consistency with dual-domain deep priors via learned transforms, soft thresholding, and exact data reinsertion. Within the learned prior, a spatial Restormer branch captures local waveform structures, while a fast Fourier convolution (FFC) branch extracts global context, effectively suppressing incoherent components and enhancing spectral energy concentration in the f−k domain. A dual-domain fusion module exchanges these complementary features via bidirectional channel attention and gated modulation, yielding substantial interpolation improvements. To overcome the severe loss of lateral continuity caused by large gaps, we introduce structural similarity (SSIM) supervision, complemented by focal frequency loss (FFL) and transform consistency constraints to suppress incoherent high-frequency components and regularize training. Experiments on a synthetic salt dome and a North Sea field dataset under 50% random and consecutive trace removal demonstrate the effectiveness of SSI-POCS. It significantly enhances event continuity across large gaps and mitigates artifacts, achieving a 3.285~dB SNR improvement over the best competing method while reducing parameters by 75.66%.
Baisong Jiang, Chunxia Zhang, Anxiang Di +4
School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, Shaanxi, 710049, China. · Geophysical Technology Research and Development Center, BGP, Zhuozhou, Hebei, 072751, China · School of Information and Communications Engineering, Xi’an Jiaotong University, Xi’an, Shaanxi, 710049, China.
Classical seismic data reconstruction relies on manually designed structural priors and iterative operators, whose coupled design space is far larger than manual trial and error can explore systematically. Deep-learning methods encode the reconstruction rules in learned weights rather than in an explicit operator that can be inspected and modified. We propose SeisEvo (Seismic Algorithm Evolution), which does not optimize a single reconstruction result but searches for the algorithm that produces it. Starting from a classical reconstruction algorithm, an LLM-driven multi-agent search modifies only the components that the user has opened for editing, without prescribing the mechanism to be discovered. Candidates that violate the physical constraints of the task are rejected outright, and the remaining ones are scored by execution. The output is neither an agent system nor a neural network, but a standalone white-box algorithm that requires no agent or neural network at inference time. For interpolation without added noise, the search discovered a residual-gated, phase-aligned dip-consistency projection; Evo-POCS improves the SNR over classic POCS by 3.49 dB on average across missing ratios from 30% to 70%. For simultaneous interpolation and denoising, it discovered a reliability-grouped singular-value shrinkage; Evo-MSSA improves the average reconstruction SNR by more than 7 dB over classic MSSA and by more than 3 dB over a stronger rank-reduction baseline. Both operators retain their gains on data not used during the search. To the best of our knowledge, this is the first study to formulate the design of a seismic reconstruction operator as a constrained, LLM-driven program evolution task. Agentic algorithm evolution can thus complement deep learning in discovering explicit, inspectable, and deployable seismic processing algorithms.
Yingjie Xu, Siwei Yu, Jianwei Ma
School of Mathematics and Center of Geophysics, Harbin Institute of Technology, Harbin, Heilongjiang, China · Department of Mathematics, National University of Singapore, Singapore · School of Earth and Space Sciences, Peking University, Beijing, China
An electromagnetic inversion scheme that integrates stochastic optimization (STO) and plug-and-play (PNP) regularization into contrast-source inversion (CSI), termed STO-PNP-CSI, is developed. Standard CSI solves for the contrast source vector of every transmitter at each iteration, which is expensive in a multi-transmitter configuration. STO instead solves for only one randomly selected contrast source vector per iteration, which reduces the per-iteration cost and can help the inversion escape poor local minima and saddle points. The resulting loss of information, however, increases the ill-posedness of the inversion. To counter this, the Swin-Conv-UNet (SCUNet) denoiser is plugged into the CSI scheme as an implicit regularizer, supplying a learned prior that is stronger than conventional hand-crafted ones and stabilizes the reconstruction. The proposed STO-PNP-CSI is applied to both synthetic and experimental data. The results show that it yields accurate reconstructions at substantially lower computational cost than CSI, including under strong nonlinearity and measurement noise.
Lingqi Gao, Hakan Bagci
Electrical and Computer Engineering (ECE) Program Computer, Electrical, and Mathematical Science and Engineering (CEMSE) Division King Abdullah University of Science and Technology (KAUST) Thuwal, 23955-6900, Saudi Arabia