Deep Learning Framework with Hybrid Loss Function for Seismic Velocity Inversion of Natural Gas Hydrate Reservoirs

Natural gas hydrate (NGH) has attracted increasing attention as a promising unconventional energy resource owing to its high volumetric storage capacity, yet its development is accompanied by significant greenhouse gas risks. Therefore, accurate NGH reservoir characterization is vital for marine resource exploration and sustainable development. Full waveform inversion (FWI) offers high-resolution velocity imaging for the NGH reservoirs, but suffers from heavy computation, sensitivity to initial models and non-uniqueness. Recent deep learning (DL) based inversion methods improve efficiency and accuracy, nevertheless, still struggle with clear boundary extraction and robustness. To overcome above difficulties, We propose SU-NeXt, a novel deep network architecture designed to learn a mapping from seismic records to velocity models, which integrates a U-Net backbone, ConvNeXt residual blocks, spatial channel squeeze and excitation attention and pixel shuffle up-sampling module. Furthermore, a hybrid loss integrating mean squared error and multi-scale structural similarity optimizes velocity value accuracy and structural fidelity. Comprehensive tests on synthetic NGH data outperforms FWI and conventional DL method in boundary delineation, structural preservation and computational efficiency. These results highlight proposed method as a reliable tool for high-resolution seismic characterization of NGH reservoirs, thereby supporting sustainable exploration and risk assessment of marine hydrate resources.