Abstract
We simulate the response of acoustic seismic waves in horizontally layered media using a deep neural network. In contrast to traditional finite-difference modelling techniques our network is able to directly approximate the recorded seismic response at multiple receiver locations in a single inference step, without needing to iteratively model the seismic wavefield through time. This results in an order of magnitude reduction in simulation time from the order of 1 s for FD modelling to the order of 0.1 s using our approach. Such a speed improvement could lead to real-time seismic simulation applications and benefit seismic inversion algorithms based on forward modelling, such as full waveform inversion. Our proof of concept deep neural network is trained using 50,000 synthetic examples of seismic waves propagating through different 2D horizontally layered velocity models. We discuss how our approach could be extended to arbitrary velocity models. Our deep neural network design is inspired by the WaveNet architecture used for speech synthesis. We also investigate using deep neural networks for simulating the full seismic wavefield and for carrying out seismic inversion directly.
Introduction: The Computational Bottleneck in Geophysics
Seismic simulation is a cornerstone of modern geophysics, essential for mapping subsurface structures in earthquake monitoring and resource exploration. Traditionally, this is achieved through Finite Difference (FD) modelling, which solves the wave equation by iteratively updating a grid over thousands of discrete time steps. While accurate, FD modelling is computationally expensive; a single 2D simulation can take seconds, and 3D volumes can take hours. This bottleneck is particularly acute in Full Waveform Inversion (FWI), where thousands of forward simulations are required to iteratively refine a geological model.
This research explores replacing these iterative solvers with deep neural networks capable of approximating wave propagation in a single inference step, potentially enabling real-time seismic analysis.
Methodology: Adapting WaveNet for Seismic Signals
The core contribution of this work is the Forward WaveNet, a model inspired by Google's WaveNet architecture for speech synthesis. The researchers treat seismic traces as time-series data, leveraging the inherent causality of wave propagation: a signal recorded at a specific time only depends on the geological layers the wave has already traversed.
The model utilizes dilated causal convolutions, allowing for a massive receptive field that captures long-term dependencies in the seismic signal without an exponential increase in parameters. The network takes a 1D velocity profile as input, converts it to a reflectivity series, and predicts the pressure response for 11 receivers simultaneously.
Beyond the forward model, the authors also investigated a Full Wavefield CNN designed to emulate a finite-difference stencil, and an Inverse WaveNet that reverses the forward process to directly predict velocity models from recorded traces.
Data and Training
To establish a reliable ground truth, the study utilized 60,000 synthetic examples (50,000 for training, 10,000 for validation) generated using the SEISMIC_CPML library. The models simulated 2D horizontally layered acoustic media with Earth-realistic distributions for layer thickness and velocity. This synthetic approach ensured that the network could be evaluated against known physical parameters.
Key Results: Speed and Fidelity
The WaveNet approach demonstrated a dramatic reduction in computation time. On a single CPU core, inference took approximately 0.18s, compared to 3.5s for the traditional FD solver. When accelerated by a GPU, inference time plummeted to 0.015s—a nearly 230x speedup over the CPU-based FD method.
Crucially, this speed does not come at the cost of accuracy for primary reflections. The model successfully learned complex physical phenomena such as Normal Moveout (NMO)—the characteristic delay in arrival times at distant sensors—and spherical spreading loss. As shown in the comparison above, the WaveNet architecture significantly outperformed a standard 1D convolutional baseline, achieving an average L2 loss of 0.03 versus the baseline's 1.7.
Limitations and Future Directions
While a successful proof of concept, the model currently faces several challenges:
- Geological Complexity: The current implementation is limited to 1D (horizontally layered) geology. Extending this to 2D and 3D structures with faults and folds is a necessary next step.
- Secondary Effects: The network struggles with "multiples"—waves that reflect multiple times between layers—which are kinematically more complex than primary reflections.
- Recursive Drift: The Full Wavefield CNN suffered from error accumulation when run over many time steps, suggesting that purely learned iterative solvers require further refinement.
Despite these limitations, the ability to approximate the wave equation's solution in a single jump represents a significant shift toward real-time geophysical imaging and more efficient inversion workflows.
Summary updated