Deep learning-based geophysical joint inversion using partial channel drop method

Joint inversion stands as a critical technique for the precise determination of subsurface structures by mitigating the ill-posedness inherent in separate geophysical inversion procedures. Recently, the integration of deep learning (DL) into joint inversion has shown promise in achieving more precise interpretations. However, existing DL-based joint inversion approaches face challenges, particularly when survey configurations between training and test datasets vary, and are prone to overfitting towards specific types of data. In response to these limitations, we introduce the Partial Channel Drop (PCD) method applied to DL joint inversion, resulting in a DL-PCD joint inversion model. Our study utilizes gravity, magnetic, and direct current resistivity data as the multiple geophysical data sources and employs 3D U-Net for the DL joint inversion model. The PCD method is implemented during the DL joint inversion training process, yielding a robust and versatile DL-based joint inversion model that can adapt to differing data configurations and manage scenarios with missing data while preventing overfitting and consequent bias in inversion results. Our proposed approach demonstrates superior generalization performance and robustness compared to separate inversion and DL joint inversion without the PCD method, exhibiting resilience even when faced with added noise. The results validate the effectiveness of the PCD method in bolstering the generalization performance of DL joint inversion, laying the groundwork for transformative possibilities in future 3D joint inversion research.