Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Equivariant Eikonal Neural Networks: Grid-Free, Scalable Travel-Time Prediction on Homogeneous Spaces
Authors: Alejandro García-Castellanos, David R. Wessels, Nicky J. van den Berg, Remco Duits, Daniel M. Pelt, Erik Bekkers
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach through applications in seismic travel-time modeling of 2D, 3D, and spherical benchmark datasets. Experimental results demonstrate superior performance, scalability, adaptability, and user controllability compared to existing Neural Operator-based Eikonal solver methods. |
| Researcher Affiliation | Collaboration | 1Amsterdam Machine Learning Lab (AMLab), University of Amsterdam 2New Theory 3Department of Mathematics and Computer Science, Eindhoven University of Technology 4Leiden Institute of Advanced Computer Science, Universiteit Leiden |
| Pseudocode | Yes | Algorithm 1 Autodecoding Training and Algorithm 2 Meta-learning Training |
| Open Source Code | Yes | Our code, including scripts to generate the results, is provided at: https://github.com/ AGarcia Cast/E-NES. |
| Open Datasets | Yes | We evaluate Equivariant Neural Eikonal Solvers (E-NES) on the 2D Open FWI benchmark [Deng et al., 2022] and extend our analysis to 3D settings to assess scalability and spherical geometry to show its generalization capabilities. Implementation details are provided in the Appendix (Section D). The code, including the experiments, is provided in the previously-mentioned public repository. The Open FWI datasets are already public at https://sites.google.com/ site/youzuolin044/openfwi. |
| Dataset Splits | Yes | For each Open FWI dataset, we sample 600 velocity fields for training and 100 for validation. We further divide the training set into 500 fields for training and 100 fields for testing. For each velocity field, we uniformly sample 20,480 coordinates, producing 10,240 pairs per velocity field. Each batch comprises two velocity fields with 5,120 source-receiver pairs per field. |
| Hardware Specification | Yes | All experiments were conducted using a single NVIDIA H100 GPU. |
| Software Dependencies | No | The paper mentions JAX implementation in the acknowledgments and also refers to a log-hyperbolic cosine loss [Jeendgar et al., 2022], but it does not provide specific version numbers for JAX or other key software libraries used for the implementation. |
| Experiment Setup | Yes | Our invariant cross-attention implementation utilizes a hidden dimension of 128 with 2 attention heads. The conditioning variables are defined as z P(Z) with cardinality |z| = 9, where Z = SE(2) R32 for each velocity field... The frequency parameters are initialized to 0.05 for the query function and 0.2 for the value function. The autodecoding phase consists of 3,000 epochs, while the meta-learning phase comprises 500 epochs. The model parameters are trained with a learning rate of 10 4. For the latent variables, context vectors use a learning rate of 10 2, while pose components in SE(2) are optimized with a learning rate of 10 3... initial learning rate of 10 4 and a minimum learning rate of 10 6. For the SGD inner loop optimization, we initialize the learning rates at 30 for context vectors and 2 for pose components, executing 5 optimization steps in the inner loop. |