PreAdaptFWI: Pretrained-Based Adaptive Residual Learning for Full-Waveform Inversion Without Dataset Dependency

TL;DR

This paper tackles the instability of Full-Waveform Inversion (FWI) under limited data and poor initial models by introducing PreAdaptFWI, a dataset-free, pretrained-based framework that couples a CNN with an Adaptive Residual Learning Module (ARLM). The method performs a dataset-free pretraining on a simple initial model to stabilize network outputs, then jointly updates the network and ARLM using FWI gradients, yielding a final velocity model that combines global stratigraphic priors with fine-grained velocity variations via $p_{ ext{Final}} = p a0 p_{ ext{Adapt}}$. Extensive experiments on Marmousi and Overthrust, including scenarios with absent low-frequency data, Gaussian noise, and extremely poor initial models, demonstrate improved convergence, accuracy (lower MAE), and robustness. The findings suggest that dataset-free pretraining plus ARLM effectively mitigates local minima and enhances practical applicability of FWI, with potential extensions to field seismic data.

Abstract

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 simple yet effective training framework that is independent of dataset reliance and requires only moderate pre-training on a simple initial model to stabilize network outputs. During the transfer learning phase, the conventional FWI gradients will simultaneously update both the neural network and the proposed adaptive residual learning module, which learns the residual mapping of large-scale distribution features in the network's output, rather than directly fitting the target mapping. Through this synergistic training paradigm, the proposed algorithm effectively infers the physically-informed prior knowledge into a global representation of stratigraphic distribution, as well as capturing subtle variations in inter-layer velocities within local details, thereby escaping local optima. Evaluating the method on two benchmark models under various conditions, including absent low-frequency data, noise interference, and differing initial models, along with corresponding ablation experiments, consistently demonstrates the superiority of the proposed approach.

Figures (18)

• Figure 1: Overview of previous FWI methods that are closely related to our approach.
• Figure 2: The schematic workflow of the proposed PreAdaptFWI. After pre-training with the observed data corresponding to a simple initial model, transfer learning is employed to achieve end-to-end prediction using real observational data.
• Figure 3: Comparison of the proposed adaptive residual learning module with conventional convolutional modules and residual learning modules.
• Figure 4: Three common-source shot gathers generated from the Marmousi model are shown. The first row represents forward data, while the second row corresponds to the same data with 10 dB additive Gaussian white noise. The introduction of noise serves to mimic the noisy conditions commonly encountered in real seismic data, enhancing the practical relevance of the study. Additionally, it provides a means to assess the inversion algorithm's robustness and its ability to withstand noise interference.
• Figure 5: Comparison of the proposed method with traditional FWI on the Marmousi model: the first row shows the initial and true models, the second row compares inversion results under normal conditions.