EikoNet: Solving the Eikonal Equation With Deep Neural Networks

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

TL;DR

EikoNet is a deep learning framework that solves the Eikonal equation—the fundamental physics of wave travel times—without relying on traditional computational grids. By training a neural network to minimize the violation of the underlying partial differential equations, the authors enable rapid, continuous, and differentiable travel-time calculations across complex 3D environments.

What Problem Does This Paper Address?

In seismology, knowing how long it takes a wave to travel from point A to point B is critical for locating earthquakes and imaging the Earth's interior (tomography). Traditionally, this is done by solving the Eikonal equation, a non-linear partial differential equation (PDE).

Standard numerical solvers like the Fast Marching Method (FMM) or Fast Sweeping Method (FSM) are "grid-based." This presents three major headaches:

  1. Memory Bloat: Storing high-resolution 3D travel-time lookup tables requires massive amounts of disk space.
  2. Interpolation Errors: If a source or receiver doesn't sit exactly on a grid node, you have to interpolate, which introduces inaccuracies.
  3. Computational Redundancy: Traditional methods solve the equation from scratch for every new source location, failing to "learn" the general physics of the velocity structure.

Methods

The authors introduce EikoNet, a mesh-independent approach based on Physics-Informed Neural Networks (PINNs).

The Core Architecture

EikoNet uses a feed-forward neural network (specifically 10 residual blocks with 512 neurons each) that takes a 6D input vector—the $(X, Y, Z)$ coordinates of both the source and the receiver. Instead of predicting travel time directly, it predicts a factored travel-time deviation ($\tau$). This factoring is a clever trick to handle the mathematical singularity (infinite gradient) that occurs exactly at the source point, which usually crashes naive neural solvers.

Training without "Labels"

Unlike standard AI that needs "ground truth" answers to learn, EikoNet is self-supervised by physics. It uses the network’s own differentiability (via backpropagation) to calculate spatial gradients. It then compares these gradients against the known velocity model using the Eikonal equation. The "loss" is simply the degree to which the network violates the laws of physics.

Smart Sampling

To train effectively, the authors explored three sampling strategies:

  • Random Distance: Uniformly sampling distances between points.
  • Random Locations: Sampling points anywhere in the volume.
  • Weighted Sampling: An importance-sampling approach where the model dynamically focuses on areas with the highest "misfit" (error), helping it learn sharp velocity contrasts more efficiently.

Datasets

The model was evaluated on several synthetic 3D velocity structures:

  • Homogeneous: Constant velocity (a baseline check).
  • Graded: Velocity increasing linearly with depth.
  • Block & Checkerboard: Models with sharp, discontinuous velocity contrasts.
  • Marmousi2: A famous, highly complex 2D geophysical benchmark representing a realistic cross-section of the Earth's crust (Angola basin).

Key Insights and Ideas

  • Continuous Domain: Because the network is a continuous function, you can query the travel time between any two coordinates. No more grid-node snapping or linear interpolation.
  • Analytical Gradients: Since the network is differentiable, you can calculate the gradient of travel time with respect to source location analytically. This is a "killer feature" for earthquake location algorithms that rely on gradient descent.
  • Transfer Learning: This is perhaps the most practical insight. Once a network learns a simple velocity model, it can be "fine-tuned" to a complex one 3–4 times faster than training from scratch. This makes it highly suitable for iterative processes like seismic tomography.
  • Scalability: The method scales independently of the number of receivers. This is vital for modern fiber-optic sensing (DAS), which creates tens of thousands of "virtual" sensors that overwhelm traditional grid-based methods.

Key Results

  • Accuracy: EikoNet matches the accuracy of the gold-standard Fast Marching Method (FMM). In the block model, it achieved a low travel-time RMS error of 0.0311s.
  • Speed: On a single NVIDIA V100 GPU, EikoNet can predict one million source-receiver travel times in just 0.424 seconds.
  • Efficiency: The entire "brain" of the 3D model (the weights) takes up only about 90MB of storage, replacing what would otherwise be gigabytes of lookup tables.

Limitations and Future Work

While EikoNet performs admirably, the authors note that in extremely complex models like Marmousi2, some "misfit" remains. The network can struggle to capture every tiny detail of high-contrast velocity structures.

Future work likely involves:

  • Developing more sophisticated neural architectures (perhaps using different activation functions or attention mechanisms).
  • Refining the weighted sampling to better handle "shadow zones" where waves don't easily travel.
  • Extending the framework to handle anisotropy (where waves travel at different speeds in different directions), which is common in the Earth's crust.

Summary updated

Abstract

The recent deep learning revolution has created enormous opportunities for accelerating compute capabilities in the context of physics-based simulations. In this article, we propose EikoNet, a deep learning approach to solving the Eikonal equation, which characterizes the first-arrival-time field in heterogeneous 3-D velocity structures. Our grid-free approach allows for rapid determination of the travel time between any two points within a continuous 3-D 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 that 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 multipathing, 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_},
}

Loading related content...

Figure