F-SiameseFWI: A Novel Deep Learning Framework for Multi-Source Full Wave Inversion

Full Wave Inversion (FWI) is an effective tool for estimating subsurface velocity models; however, its high nonlinearity presents certain limitations, including significant computational costs in terms of time and hardware resources. Accelerating the FWI inversion process without compromising performance remains a challenging task. To address this issue, we propose the FreqSiameseFWI framework, a deep learning-based misfit function designed to expedite the FWI inversion process and support Multi-source FWI (MSFWI) by mitigating the impact of cross-talk noise. FreqSiameseFWI employs a self-supervised learning approach integrated within the FWI framework, enabling iterative updates of its parameters without introducing significant additional overhead costs. This method is grounded in the Siamese network architecture, which facilitates comparisons between input seismic data in their latent representation, thereby enhancing the fidelity of these comparisons. The seismic data are converted to the frequency domain through Fast Fourier Transform (FFT), allowing the Siamese network to extract spectral features from the seismic data and achieve robust inversion performance that is beneficial for both FWI and MSFWI applications. The proposed FreqSiameseFWI markedly accelerates the FWI inversion process, producing speed improvements that are proportional to the number of stacked sources used to generate super-shots, while effectively mitigating the influence of cross-talk noise. The performance of FreqSiameseFWI, evaluated using the Marmousi2 and Overthurst models, has yielded promising results that outperform conventional misfit functions.