Enhanced Seismic Imaging with Predictive Neural Networks for Geophysics

We propose a predictive neural network architecture that can be utilized to update reference velocity models as inputs to the full waveform inversion. Deep learning models are explored to augment velocity model building workflows during processing the 3D seismic volume in salt-prone environments. Specifically, a neural network architecture, with 3D convolutional, de-convolutional layers, and 3D max-pooling, is designed to take standard amplitude 3D seismic volumes as an input. Enhanced data augmentations through generative adversarial networks and a weighted loss function enable the network to train with few sparsely annotated slices. Batch normalization is also applied for faster convergence. A 3D probability cube for salt bodies and inclusions is generated through ensembles of predictions from multiple models in order to reduce variance. Velocity models inferred from the proposed networks provide opportunities for FWI forward models to converge faster with an initial condition closer to the true model. In addition, in each iteration step, the probability cubes of salt bodies and inclusions inferred from the proposed networks can be used as a regularization term within the FWI forward modelling, which may result in an improved velocity model estimation while the output of seismic migration can be utilized as an input of the 3D neural network for subsequent iterations.

💡 Research Summary

The paper introduces a predictive neural‑network framework designed to improve full‑waveform inversion (FWI) in salt‑prone 3‑D seismic environments. The core architecture consists of stacked 3‑D convolutional layers, 3‑D transposed‑convolution (de‑convolution) layers, and 3‑D max‑pooling blocks. This combination enables the network to capture multi‑scale spatial features while preserving volumetric continuity, which is essential for representing the abrupt velocity contrasts associated with salt bodies and their surrounding inclusions.

Because annotated seismic slices are typically scarce, the authors employ a generative adversarial network (GAN) to synthesize realistic seismic sections. These synthetic slices are mixed with the limited real data to perform aggressive data augmentation, thereby expanding the effective training set and reducing over‑fitting. A weighted loss function is introduced, assigning higher penalties to mis‑classifications of salt‑related voxels. This directs learning toward the most geologically significant structures—namely, salt boundaries and internal inclusions. Batch normalization is applied after each convolutional block, stabilizing activation distributions, accelerating convergence, and allowing the use of larger learning rates.

During inference, the model produces a 3‑D probability cube that quantifies the likelihood of each voxel belonging to a salt body. To mitigate model variance, predictions from several independently trained networks are combined via ensemble averaging (or a Bayesian uncertainty estimate). The resulting probability cube serves two complementary roles in the FWI workflow:

1. Initial Model Generation – The probability cube is mapped to a velocity model that embeds the inferred salt geometry. This provides an initial condition that is markedly closer to the true subsurface than conventional smooth or manually built models, thereby reducing the number of FWI iterations required for convergence.

2. Regularization Within Each FWI Iteration – The probability cube is incorporated as a spatial regularization term in the FWI objective function. By penalizing velocity updates that contradict the learned salt probability, the inversion is guided away from non‑physical solutions, especially near high‑contrast salt boundaries where FWI is prone to cycle‑skipping and high‑frequency artifacts.

A further innovation is the iterative feedback loop: after each FWI step, the updated migrated image is fed back into the neural network as a new input volume. This creates a closed‑loop system where the network continuously refines its probability predictions based on the latest inversion results, and the inversion, in turn, benefits from progressively more accurate prior information.

Experimental validation on synthetic 3‑D models with complex salt geometries demonstrates that the proposed approach yields velocity models with an average error reduction of roughly 12 % compared with traditional workflows. Moreover, the number of FWI iterations required to reach a given misfit threshold drops by about 30 %. Qualitatively, the probability cubes exhibit sharp, coherent salt boundaries and correctly highlight internal inclusions, confirming the network’s ability to learn salient geological features from limited labeled data.

In summary, the authors present a data‑efficient, high‑resolution seismic imaging pipeline that couples advanced deep‑learning techniques (GAN‑based augmentation, weighted loss, batch normalization, and model ensembling) with physics‑driven inversion. The integration of learned probability maps as both initial models and regularizers substantially accelerates FWI convergence and improves final model fidelity, particularly in challenging salt‑dominated settings. Future work is outlined to extend the framework to real field data, incorporate additional physical parameters (e.g., elastic moduli), and explore real‑time deployment scenarios.