Cyclic Learning Rate U-Shaped ResNet Embedded with Dual Attentions for Velocity Model Building

Data-driven full-waveform inversion shows great potential for subsurface velocity estimation, but its accuracy is often hindered by susceptibility to local minima and the inherent characteristics of seismic data. Seismic shot gathers typically exhibit weak spatial correlations between data and velocity models, further complicated by sparsity due to reflected events. This study proposes a cyclic learning rate velocity model building (CLR-VMB) framework, incorporating a U-shaped residual network (ResNet) enhanced with dual attention mechanisms (DA-ResNet) to address these challenges. Unlike conventional single-network approaches, CLR-VMB leverages multiple snapshots of the network obtained during a cyclic learning rate schedule. This enables the framework to ensemble complementary information from different snapshots without additional training costs. The DA-ResNet architecture utilizes attention mechanisms to extract cross-gather relationships and capture crucial spatial information from seismic shot gathers. An encoder-decoder structure is then used to estimate velocity from the processed recordings. Numerical experiments conducted on various layered models and SEG/EAGE salt models demonstrate that CLR-VMB effectively mitigates local minima, enhances noise suppression, and exhibits superior generalization capabilities compared to a UNet-based method and a physics-guided hybrid approach. By intelligently interpreting inherent seismic data characteristics, this study's cyclic learning and attention-based framework paves the way for effective data-driven approaches to subsurface velocity building.