<p>The direct current (DC) resistivity method is extensively employed in the investigation of challenging geological conditions. The inversion of resistivity model based on observed data represents a prevalent approach for geological interpretation. In recent years, significant strides have been made in this field through the application of deep learning techniques. Unsupervised learning methods that incorporate physics principles offer particular promise due to their independence from manual annotation. However, gradients derived from electric field propagation physics rules, often suffer from low quality and high computational complexity, thereby yielding suboptimal predictive results. In our study, we propose a gradient optimization approach for unsupervised deep learning (DL) inversion. Firstly, we perform multiple gradient calculations and aggregate them, thereby updating the resistivity model with the cumulative gradients to mitigate errors. Secondly, we introduce a novel method for rapidly solving sensitivity matrices by optimizing matrix computations and data storage, resulting in a significant enhancement in computational efficiency by several orders of magnitude. Finally, we validate the effectiveness of our gradient optimization approach through comprehensive experiments and field tests, accurately locating and depicting both boulder and karst zones.</p>