A new way to map underground properties using reinforcement learning
DeepWaveRL uses a reward system to estimate subsurface velocity from seismic measurements.
Full Waveform Inversion (FWI) is a fundamental inverse problem. It estimates subsurface properties like velocity from seismic measurements.
DeepWaveRL incorporates reinforcement learning into self-supervised FWI. This framework reformulates FWI as a policy learning problem.
A policy network predicts velocity maps as actions. A forward operator is used solely for simulation.
A global scalar reward is computed based on data misfit for each prediction. This reward weights policy gradients.
Learning proceeds without accessing the gradients from the forward operator. This happens in a challenging setting without accessing those gradients.
The findings suggest that this framework offers a flexible and complementary learning paradigm for the FWI research community.
Key facts
- DeepWaveRL achieves competitive performance compared to supervised and self-supervised baselines.
- The framework uses a global scalar reward based on data misfit.
- The method works without accessing gradients from the forward operator.