SiameseFit: Latent Seismic Data Comparison for Enhanced Full Waveform Inversion and Cycle‐Skipping Mitigation

The proper comparison between observed and simulated data is at the core of achieving robust inversion results in high‐dimensional optimization problems, like full waveform inversion (FWI). However, this comparison is challenging when the initial model for FWI is poor, resulting in simulated seismic data that is far away (beyond a half cycle) from the observed data. To address this issue, we propose a self‐supervised deep learning‐based misfit function, termed SiameseFit , designed to mitigate the cycle‐skipping problem. The core of the proposed SiameseFit framework is a Siamese neural network consisting of two identical convolutional neural network branches with shared weights, which extract features from the observed and simulated data described by a reduced‐dimensional latent space. The network parameters are randomly initialized and jointly optimized during the FWI process. The comparison between the observed and simulated data is performed in the latent space using the Euclidean distance, enabling more robust feature alignment and improved inversion performance. The proposed SiameseFit framework for FWI demonstrates robust performance in mitigating cycle skipping when applied to two synthetic examples and one field dataset, outperforming both traditional FWI misfit functions and benchmark deep learning models. Furthermore, SiameseFit‐based FWI effectively suppresses crosstalk noise in multi‐source implementations, achieves convergence to a high‐resolution model within only 50 FWI (Hessian‐free) iterations when provided with a good initial model and is capable of converging even from a constant velocity model. Our analysis further demonstrates that the latent representation learned by SiameseFit effectively corrects misalignments between the observed and simulated data in the latent space, leading to improved adjoint source updates and more accurate velocity model reconstruction.