AI
New framework unites geometry and neural networks for faster travel-time predictions
Equivariant Eikonal Neural Networks enable grid-free computation of travel times across curved spaces, with applications to seismic modeling.
A team of researchers has proposed a hybrid approach that combines equivariant neural fields with physics-informed neural networks to tackle the challenge of predicting travel times across complex geometric spaces without relying on computational grids.
The core innovation#
The method, called Equivariant Neural Eikonal Solvers, leverages a single neural field architecture where a shared backbone adapts to different problem instances through latent variables encoded as point clouds in a mathematical structure called a Lie group. This design delivers weight-sharing efficiency, ensures solutions respect the underlying geometry, and critically, enables steerability—users can transform the latent representation to induce predictable, geometrically consistent changes in the output.
What makes it work#
By grounding equivariant mappings into physics-informed training, the framework learns to solve the Eikonal equation (which governs wave and travel-time propagation) across diverse curved manifolds: Euclidean space, position-orientation configurations, spheres, and hyperbolic geometries. The approach sidesteps the need for fixed computational grids, a constraint that limits many existing methods.
Validation and performance#
Testing on seismic travel-time benchmarks in 2D, 3D, and spherical settings, the framework outperformed comparable Neural Operator-based Eikonal solvers in speed, scalability, adaptability, and user control. The results suggest the technique could improve modeling in geophysics and other domains where understanding how signals propagate through complex media is essential.
