EikoNet: Solving the Eikonal equation with Deep Neural Networks

Jonathan D. Smith, Kamyar Azizzadenesheli, Zachary E. Ross

Abstract

The recent deep learning revolution has created an enormous opportunity for accelerating compute capabilities in the context of physics-based simulations. Here, we propose EikoNet, a deep learning approach to solving the Eikonal equation, which characterizes the first-arrival-time field in heterogeneous 3D velocity structures. Our grid-free approach allows for rapid determination of the travel time between any two points within a continuous 3D domain. These travel time solutions are allowed to violate the differential equation - which casts the problem as one of optimization - with the goal of finding network parameters that minimize the degree to which the equation is violated. In doing so, the method exploits the differentiability of neural networks to calculate the spatial gradients analytically, meaning the network can be trained on its own without ever needing solutions from a finite difference algorithm. EikoNet is rigorously tested on several velocity models and sampling methods to demonstrate robustness and versatility. Training and inference are highly parallelized, making the approach well-suited for GPUs. EikoNet has low memory overhead, and further avoids the need for travel-time lookup tables. The developed approach has important applications to earthquake hypocenter inversion, ray multi-pathing, and tomographic modeling, as well as to other fields beyond seismology where ray tracing is essential.

Introduction

In seismology, calculating the travel time of waves between a source and a receiver is fundamental for earthquake localization and seismic tomography. Traditionally, this involves solving the Eikonal equation, a non-linear partial differential equation (PDE). Conventional numerical solvers, such as the Fast Marching Method (FMM) or Fast Sweeping Method (FSM), rely on discrete computational grids. These grid-based approaches suffer from significant drawbacks: high memory consumption for 3D lookup tables, interpolation errors when points do not align with grid nodes, and the need to re-solve the entire domain for every new source location.

EikoNet addresses these challenges by introducing a grid-free, physics-informed deep learning framework. By leveraging the differentiability of neural networks, EikoNet provides a continuous and differentiable solution to the Eikonal equation across complex 3D velocity structures.

Methodology

Network Architecture and Formulation

EikoNet utilizes a feed-forward neural network consisting of 10 residual blocks, each containing 512 neurons. The network accepts a 6D input vector representing the $(X, Y, Z)$ coordinates of both the source ($\vec{X}_s$) and the receiver ($\vec{X}_r$). To avoid mathematical singularities at the source point, the network is designed to predict a factored travel-time deviation ($\tau$).

Overview of the EikoNet framework, including network architecture, physics-informed formulation, sampling strategies, and training performance.

Physics-Informed Training

Unlike traditional supervised learning, EikoNet is trained using a self-supervised, physics-informed approach. The loss function is defined by the degree to which the network's predictions violate the Eikonal equation:

$$\|\nabla_r T_{s \to r}\|^2 = \frac{1}{V_r^2}$$

Spatial gradients are calculated analytically through the network using backpropagation. This allows the model to learn the underlying physics directly from the velocity model without requiring pre-computed ground-truth travel times.

Sampling Strategies

To ensure robust learning across the 3D domain, the authors implemented several sampling strategies:

  • Random Distance/Locations: Uniform sampling to cover the volume.
  • Weighted Sampling: An importance-sampling technique that focuses training on regions with higher "misfit" or error, facilitating the learning of sharp velocity contrasts.

Results and Discussion

Performance and Accuracy

EikoNet demonstrates accuracy comparable to the gold-standard FMM. In a block velocity model, it achieved a travel-time RMS error of 0.0311s. The method is highly efficient; on an NVIDIA V100 GPU, EikoNet can compute one million source-receiver travel times in approximately 0.424 seconds.

Efficiency and Scalability

A significant advantage of EikoNet is its low memory footprint. The entire trained model (approximately 90MB) replaces gigabytes of traditional lookup tables. Furthermore, because the network is a continuous function, it eliminates the need for interpolation and scales efficiently with the number of receivers, making it ideal for high-density sensing technologies like Distributed Acoustic Sensing (DAS).

Transfer Learning

The framework supports transfer learning, where a model pre-trained on a simple velocity structure can be fine-tuned for a complex one 3–4 times faster than starting from scratch. This capability is particularly valuable for iterative geophysical inversion processes.

Conclusion

EikoNet represents a shift from discrete grid-based solvers to continuous, differentiable neural representations for seismic ray tracing. While challenges remain in capturing extremely fine details in high-contrast models like Marmousi2, the method's speed, memory efficiency, and analytical differentiability offer a powerful new tool for modern seismology and beyond.

Summary updated

Abstract

The recent deep learning revolution has created an enormous opportunity for accelerating compute capabilities in the context of physics-based simulations. Here, we propose EikoNet, a deep learning approach to solving the Eikonal equation, which characterizes the first-arrival-time field in heterogeneous 3D velocity structures. Our grid-free approach allows for rapid determination of the travel time between any two points within a continuous 3D domain. These travel time solutions are allowed to violate the differential equation - which casts the problem as one of optimization - with the goal of finding network parameters that minimize the degree to which the equation is violated. In doing so, the method exploits the differentiability of neural networks to calculate the spatial gradients analytically, meaning the network can be trained on its own without ever needing solutions from a finite difference algorithm. EikoNet is rigorously tested on several velocity models and sampling methods to demonstrate robustness and versatility. Training and inference are highly parallelized, making the approach well-suited for GPUs. EikoNet has low memory overhead, and further avoids the need for travel-time lookup tables. The developed approach has important applications to earthquake hypocenter inversion, ray multi-pathing, and tomographic modeling, as well as to other fields beyond seismology where ray tracing is essential.
BibTeX
@article{smith2021eikonet,
  author = {Jonathan D. Smith and Kamyar Azizzadenesheli and Zachary E. Ross},
  title = {{EikoNet: Solving the Eikonal equation with Deep Neural Networks}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2021},
  url = {https://openreview.net/forum?id=UWjKwtqPQ_},
}
Figure