Optimizing Regularization Parameters for Potential Field Data Inversion Using RPSNN

An important approach to obtain the underground structure is the inversion of potential field data, which is an ill-posed problem. Regularization inversion is a commonly used method for addressing such ill-posed problems. The selection of regularization parameters determines the accuracy and stability of the inversion. However, conventional methods are computationally complex, and the calculation process may introduce errors in parameter selection. We discuss the shortcomings of existing methods and propose a regularization parameter selection neural network (RPSNN) to calculate regularization parameters for potential field data inversion. This network is designed to more effectively capture the relationship between potential field data and regularization parameters. Comparative analysis reveals that our method can directly obtain the optimal regularization parameter (ORP) and lead to more accurate inversion results. Therefore, our method is more effective in selecting regularization parameters for potential field data inversion. Numerical experiments demonstrate that trained RPSNN outperforms existing methods in predicting regularization parameters, leading to improved solutions for inverse problems. The proposed method was applied to the Lu-Zong ore concentration area in the middle and lower reaches of the Yangtze River region to obtain a 3-D distribution model of magnetite within 5 km underground. The results showed that the magnetite locations corresponded well with the shallow minerals being mined and were verified with drilling data, confirming the existence of usable deep resources in the ore concentration area. The extent of deep minerals was circled to delineate favorable mineralization areas.