DispSAC: Surface-wave dispersion inversion using soft actor-critic reinforcement learning
Abstract
Surface-wave inversion (SWI) is widely used to image shear-wave velocity structure in the crust and upper mantle. Locally linearized methods offer computational efficiency but can be sensitive to the initial model and regularization, whereas Bayesian sampling methods can characterize model uncertainty at substantial computational cost. Neural networks provide another route to efficient inversion, although deterministic predictions alone do not describe the range of models compatible with the data. We present DispSAC, an iterative SWI method based on the off-policy soft actor-critic reinforcement-learning algorithm. The agent is trained using synthetic Rayleigh-wave dispersion curves generated from isotropic shear-wave velocity models parameterized by B-splines. During inference, the trained policy repeatedly updates the model parameters using feedback from a physical forward solver. Multiple initializations produce an ensemble of recovered models whose spread describes variability among the solutions obtained. We evaluate DispSAC using synthetic test data drawn from the training distribution, synthetic data derived from a western United States reference model, and observed dispersion data from Yunnan, China. The results demonstrate that the learned policy can reduce dispersion misfit across these datasets and recover regional velocity structure. Comparisons with the implemented Markov chain Monte Carlo baseline suggest potential computational benefits under the tested conditions. DispSAC offers a sufficiently generalized framework for reusing inversion experience across locations and for future integration with additional geophysical observations.
Related
- DeepWaveRL: Self-Supervised Full Waveform Inversion via Reinforcement Learning
- Multi-Modal Surface Wave Inversion using Physics Guided Variational Autoencoders
- EnviFlux (v1.0): a simplified surface flux inversion tool based on four-dimensional variational data assimilation (4D-Var)
- Meta reinforcement learning for robust quadruped locomotion under joint failures
- Robust reinforcement learning multi-baseline phase unwrapping algorithm with JPEG-LS
BibTeX
@misc{tian2026dispsac,
title = {{DispSAC: Surface-wave dispersion inversion using soft actor-critic reinforcement learning}},
author = {Mingteng Tian and Yingjie Yang},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5mj75},
url = {https://eartharxiv.org/repository/view/15137/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - DispSAC: Surface-wave dispersion inversion using soft actor-critic reinforcement learning
AU - Mingteng Tian
AU - Yingjie Yang
PY - 2026
DA - 2026/09/23
PB - EarthArXiv
DO - 10.31223/x5mj75
UR - https://eartharxiv.org/repository/view/15137/
AB - Surface-wave inversion (SWI) is widely used to image shear-wave velocity structure in the crust and upper mantle. Locally linearized methods offer computational efficiency but can be sensitive to the initial model and regularization, whereas Bayesian sampling methods can characterize model uncertainty at substantial computational cost. Neural networks provide another route to efficient inversion, although deterministic predictions alone do not describe the range of models compatible with the data. We present DispSAC, an iterative SWI method based on the off-policy soft actor-critic reinforcement-learning algorithm. The agent is trained using synthetic Rayleigh-wave dispersion curves generated from isotropic shear-wave velocity models parameterized by B-splines. During inference, the trained policy repeatedly updates the model parameters using feedback from a physical forward solver. Multiple initializations produce an ensemble of recovered models whose spread describes variability among the solutions obtained. We evaluate DispSAC using synthetic test data drawn from the training distribution, synthetic data derived from a western United States reference model, and observed dispersion data from Yunnan, China. The results demonstrate that the learned policy can reduce dispersion misfit across these datasets and recover regional velocity structure. Comparisons with the implemented Markov chain Monte Carlo baseline suggest potential computational benefits under the tested conditions. DispSAC offers a sufficiently generalized framework for reusing inversion experience across locations and for future integration with additional geophysical observations.
ER -