ContextSeisNet: In-Context Seismic Demultiple

• ContextSeisNet is a deep learning framework for seismic demultiple that uses in-context learning to condition predictions on neighboring gathers for spatially consistent results.
• It integrates support seismic gathers and corresponding ground-truth results to enforce lateral continuity and offer user control over processing characteristics.
• The framework achieves up to +2–4 dB PSNR gain with a 90% reduction in required training data compared to traditional U-Net architectures.

ContextSeisNet is a deep learning framework for seismic data processing, specifically designed to leverage in-context learning (ICL) for spatially consistent demultiple—a critical pre-stack task in seismic imaging. Unlike traditional demultiple algorithms or standard supervised neural networks, ContextSeisNet conditions its predictions on a support set of neighboring seismic gathers and their corresponding ground-truth demultiple results. This formulation enables the model to adapt task-specific behavior at inference time, enforce lateral coherence, and offer user control over processing characteristics, all while requiring substantially fewer labeled training examples than conventional approaches (Fuchs et al., 12 Dec 2025).

1. Seismic Demultiple and the Motivation for In-Context Learning

Seismic demultiple involves suppressing multiple reflections—energy that traverses the subsurface more than once—while preserving primary reflections. These multiples, if not removed, obscure subsurface imaging and complicate stratigraphic interpretation. Classical demultiple techniques such as Surface-Related Multiple Elimination (SRME) and Radon-transform-based muting depend on dense acquisition geometry, manual parameterization, and typically process each common-depth-point (CDP) gather independently, resulting in artifacts and lateral inconsistencies.

Conventional deep learning, typified by applying U-Nets to individual gathers, similarly treats each CDP as an isolated task, failing to guarantee spatial coherence and offering no recourse for user-directed inference. In contrast, the ICL paradigm allows query predictions to be conditioned on user-supplied prompts—neighboring CDPs and their labels—which can be chosen from processed field data or provided interactively.

2. ContextSeisNet Model Architecture

ContextSeisNet adapts the UniverSeg segmentation framework to seismic demultiple. The high-level architecture is a U-Net-style encoder-decoder network where each block is supplanted by a CrossBlock that fuses features from the query and each support example, and aggregates these via permutation-invariant averaging:

• Encoder: Processes the query gather in conjunction with S support gathers, each through identical encoder pathways composed of CrossBlocks. At each stage, the query and all supports are merged using channel-wise concatenation and processed by shared convolutional kernels.
• CrossBlock: At level $l$, let $u \in \mathbb{R}^{C \times H \times W}$ be the query feature map, and $\{V^s_{\text{feature}}\}_{s=1}^S$ the S support feature maps. For each support $s$, compute

$$z_s = \text{Conv}_{\theta_{C1}}(u \| V^s_{\text{feature}})$$

(where $\|$ denotes concatenation). Each $z_s$ is further processed and nonlinearly transformed, and the query feature is updated based on the average over all supports:

$$u' = \sigma\left(\text{Norm}\left(\text{Conv}_{\theta_{C3}} \left( \frac{1}{S} \sum_{s=1}^{S} z_s \right)\right)\right)$$

where $\sigma$ is LeakyReLU and Norm is batch normalization. This ensures that feature adaptation is both local (via matching to similar gathers) and spatially smooth.

• Decoder: Mirrors the encoder, utilizing skip connections and additional CrossBlocks to propagate and merge context-aware multiscale features for final prediction.

3. In-Context Learning Formulation

The formal mapping realized by ContextSeisNet is

$f^{\text{ICL}}_\theta(X_q \mid V) = Y_q$

where $X_q$ is the query CDP gather, $V = \{(X_s, Y_s)\}_{s=1}^S$ is the support set (input-label pairs), and $Y_q$ is the predicted primary-only gather. During training, batches are sampled as a tuple: one query, S supports from the same line (possibly with random replacement), and all undergo data augmentation (noise and normalization).

The objective is to minimize the L1 loss:

$$\ell(\theta) = L_1(f^{\text{ICL}}_\theta(X \mid V), Y)$$

with stochastic gradient descent (AdamW, OneCycle schedule). The inclusion of gather–label pairs as support at inference enables dynamic, user-steered and context-dependent processing.

4. Spatial Consistency and User Control Mechanisms

Spatial consistency is attained by including neighboring gathers in the prompt set. The CrossBlock mechanism ensures that the query output is "attracted" toward representations that match the average support, reinforcing smooth transitions and discouraging abrupt event type changes between adjacent gathers. This feature-level regularization is realized algorithmically: every update to the query feature $u'$ explicitly blends local support-based statistics.

User control arises from the flexibility in defining the prompt set V: experts can specify support examples that encode the desired processing aggressiveness (e.g., mild or strong multiple suppression), or use outputs of conventional or learned demultiple algorithms as in-context anchors. This enables domain adaptation and the propagation of field-specific processing styles without retraining.

5. Training Regimes, Hyperparameters, and Data Efficiency

Training leverages large-scale synthetic datasets: e.g., 15,000 seismic lines, each comprising 21 spatially correlated CDPs with 64 offsets × 256 time samples. Primary and multiple events are generated by convolutional modeling with laterally varying normal moveout (NMO) velocities. Support set size during training is varied (1, 3, 5, 10) to study its effect.

• Data augmentations: random white noise addition, per-gather normalization, and random identity replacement in labels.
• Optimizer: AdamW (lr=0.001, weight decay=0.01), training batch size=64 (1 query + S support), gradient clipping to norm 1.
• ContextSeisNet demonstrates high data efficiency: on held-out field data, comparable performance to a U-Net trained on 100,000 labeled gathers is achieved using only 10,500 synthetic training examples, representing a 90% reduction (Fuchs et al., 12 Dec 2025).

6. Quantitative Evaluation and Ablation Studies

Benchmarks show ContextSeisNet outperforms U-Net baselines in both PSNR and lateral coherence:

Model/Setup	PSNR Gain (synthetic)	Lateral Coherence	Data Efficiency
U-Net (100k)	--	Flicker	Baseline
U-Net (10k)	--	Poor	--
ContextSeisNet (10k, ICL)	+2–4 dB	Superior	90% reduction vs U-Net

Longer support sets (S=5 or 10) extend high-quality predictions farther from prompt positions, while prompt spacing impacts uniformity of performance along the survey line. Ablations reveal that increasing S enhances local PSNR near prompts but can introduce marginal quality degradation at queries very near prompt positions if over-parameterized. Broader prompt spacing yields nearly uniform performance across the entire line.

On field data, ContextSeisNet eliminates lateral flicker and improves near-offset multiple removal, inheriting the desired characteristics of the chosen prompt set (e.g., Radon-based demultiple). No p-values are reported, but improvements are visually and quantitatively documented (Fuchs et al., 12 Dec 2025).

7. Limitations, Extensions, and Integration with Contextual Architectures

Limitations include the requirement for fixed-size support sets (static computational graphs), and dependence on prompt quality—aggressive prompts risk over-subtraction; poor prompts risk event retention. Error propagation is possible with sequential re-prompting. Potential extensions include application to prestack tasks (alignment, denoising), post-stack interpretation (fault detection, salt-body picking), and integration of multimodal support (e.g., GAN or diffusion model outputs as prompts).

From the spatial context-injection strategy developed for seismic inversion using 2D Temporal Convolutional Networks (TCNs) (Mustafa et al., 2021), several additional architectural enhancements for ContextSeisNet are suggested:

• Extension of contextual blocks to 2D spatio-temporal operations with dilated convolutions.
• Incorporation of auxiliary reconstruction objectives (multitask learning) for robustness and regularization.
• Weight sharing with joint training across surveys, implemented as a soft L2 penalty or analogous dual-path architectures, to leverage domain-general features without risking negative transfer.
• Hyperparameter tuning (such as the coupling weight α in TCN sharing) to balance shared versus dataset-specific representation learning.

These architectural motifs are expected to further reinforce lateral continuity, noise robustness, and overfitting resistance in data-scarce settings.

• "In-Context Learning for Seismic Data Processing" (Fuchs et al., 12 Dec 2025)
• "Joint Learning for Spatial Context-based Seismic Inversion of Multiple Datasets for Improved Generalizability and Robustness" (Mustafa et al., 2021)
• For contextual comparison: "Seismic-Net: A Deep Densely Connected Neural Network to Detect Seismic Events" (Wu et al., 2018)