Bidirectional Physics-Constrained Full Waveform Inversion: Reducing Seismic Data Dependency in Velocity Model Building

Summary Full-waveform inversion (FWI) is a method that utilizes seismic data to invert the physical parameters of subsurface media by minimizing the difference between simulated and observed waveforms. Due to its ill-posed nature, FWI is susceptible to getting trapped in local minima. Consequently, various research efforts have attempted to combine neural networks with FWI to stabilize the inversion process. This study presents a bidirectional physics-constrained full waveform inversion (BP-FWI) framework that leverages transfer learning by pre-training on simple initial models and utilizing the results. Additionally, it employs FWI gradients to co-optimize both the neural network and the adaptive residual learning module under bidirectional physics constraints. By eliminating the reliance on a large amount of manually constructed synthetic datasets, the proposed training strategy addresses the challenge of data dependency. Furthermore, through the joint optimization strategy guided by bidirectional constraints, the neural network is able to focus on integrating physically-informed prior knowledge into global stratigraphic representations, while the adaptive residual learning module specializes in learning residual mappings from the network’s output, thereby capturing subtle inter-layer velocity variations in local geological structures. Evaluating the method on two benchmark models under various conditions, including absent low-frequency data, noise interference, non-uniform receiver configurations, and differing initial models, along with corresponding ablation experiments, consistently demonstrates the superiority of the proposed approach.