Coarse-to-fine multi-resolution hash encoding for implicit full waveform inversion
Abstract
Implicit full waveform inversion (IFWI) reparameterizes the velocity model with a neural network, but it suffers from spectral bias and requires many iterations to converge. Multi-resolution hash encoding uses a learnable feature grid to recover high wavenumber details at multiscales and provides strong local representations. However, activating all resolution levels from the start of the inversion introduces high wavenumber components, which can exacerbate cycle skipping. In this paper, we propose a coarse-to-fine IFWI (C2F-IFWI) that activates the multi-resolution hash levels progressively rather than all at once. Specifically, only a few coarse base levels are kept active at initialization, while the remaining levels are introduced progressively. A frozen hash level contributes no gradient until it is activated, so the coarse levels resolve the long-wavelength background before the finer levels are introduced regardless of the frequency range of the data. A continuous activation index increases with iterations and determines how many levels are active at any point in training. Each newly admitted level is then faded in gradually through a smooth weighting rather than switched on abruptly, which keeps the optimization stable during transition. This scheme allows model capacity to grow synchronously with wavenumber without adding any additional trainable parameters and with negligible computational overhead. Experiments on the synthetic Overthrust and BP 2004 models and on the Viking marine field dataset show that, compared with the all-resolution hash baseline, the coarse-to-fine scheme converges faster, attains higher accuracy, and is markedly more robust to the choice of initial model and to strong data noise.
Related
- Seismic imaging of the crust and upper mantle beneath Taiwan using full-waveform inversion of teleseismic body waves
- A comprehensive review of strategies to mitigate non-convexity in full waveform inversion
- Uncertainty Quantification in Full Waveform Inversion with Gradient Flows in Probability Space
- Effect of the interfacial continuity on the elastic full-waveform inversion
- Full Waveform Inversion with a Learned Preconditioner
BibTeX
@misc{wang2026coarsetofine,
title = {{Coarse-to-fine multi-resolution hash encoding for implicit full waveform inversion}},
author = {Linrong Wang and Shaowen Wang and Fan Min and Tariq Alkhalifah},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.01081},
eprint = {2610.01081},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.01081},
}
RIS
TY - UNPB
TI - Coarse-to-fine multi-resolution hash encoding for implicit full waveform inversion
AU - Linrong Wang
AU - Shaowen Wang
AU - Fan Min
AU - Tariq Alkhalifah
PY - 2026
DA - 2026/10/01
PB - arXiv
DO - 10.48550/arxiv.2610.01081
UR - https://arxiv.org/abs/2610.01081
AB - Implicit full waveform inversion (IFWI) reparameterizes the velocity model with a neural network, but it suffers from spectral bias and requires many iterations to converge. Multi-resolution hash encoding uses a learnable feature grid to recover high wavenumber details at multiscales and provides strong local representations. However, activating all resolution levels from the start of the inversion introduces high wavenumber components, which can exacerbate cycle skipping. In this paper, we propose a coarse-to-fine IFWI (C2F-IFWI) that activates the multi-resolution hash levels progressively rather than all at once. Specifically, only a few coarse base levels are kept active at initialization, while the remaining levels are introduced progressively. A frozen hash level contributes no gradient until it is activated, so the coarse levels resolve the long-wavelength background before the finer levels are introduced regardless of the frequency range of the data. A continuous activation index increases with iterations and determines how many levels are active at any point in training. Each newly admitted level is then faded in gradually through a smooth weighting rather than switched on abruptly, which keeps the optimization stable during transition. This scheme allows model capacity to grow synchronously with wavenumber without adding any additional trainable parameters and with negligible computational overhead. Experiments on the synthetic Overthrust and BP 2004 models and on the Viking marine field dataset show that, compared with the all-resolution hash baseline, the coarse-to-fine scheme converges faster, attains higher accuracy, and is markedly more robust to the choice of initial model and to strong data noise.
ER -