Abstract: The short-offset transient electromagnetic (SOTEM) data are typically processed using conventional inversion methods based on physical modeling, manifesting relatively low efficiency and difficulty in integrating priori information. In contrast, the data-driven inversion methods can enhance the inversion accuracy and efficiency but fail to ensure the generalization capability. To achieve high inversion accuracy and efficiency for SOTEM data and a reliable generalization capability, this study proposed an inversion method that integrates physical modeling with the data-driven approach, introducing the supervised descent method in machine learning into SOTEM data inversion. The proposed inversion method involves the offline training and online prediction stages. In the offline training stage, the prior information is flexibly integrated into the model training through a reasonable training dataset to obtain the average descent directions with implicit model features. In the online prediction stage, the physical modeling functions and the descent directions are employed to reconstruct the model parameters under the conventional inversion framework. In this study, the layered geodetic model was applied to design the training and test datasets for the 1D inversion of SOTEM data based on the supervised descent method. The inversion results were compared with those obtained using Occam's inversion algorithm, demonstrating that the proposed inversion method shows significantly enhanced inversion efficiency, higher inversion accuracy, and higher generalization capability.