Fast Inversion of Subsurface Target Electromagnetic Induction Response With Deep Learning

With the development of electromagnetic detection technology, the inversion of electromagnetic induction (EMI) response based on a 3-D orthogonal dipole model can provide an estimation of parameters of a high-conductivity target, such as the target's position, orientation, and shape. However, the traditional inversion methods suffer from several limitations including relying on the initial value, easy to fall into the local optimal solution, and high computational complexity. To overcome these disadvantages, in this letter, we propose a deep learning (DL) inversion method of subsurface target EMI response based on deep neural network (DNN) architecture, which is combined with adaptive moment estimation (Adam) optimization algorithm and learning rate attenuation strategy to improve the model accuracy. Datasets are obtained from forward modeling in different target parameters. By limiting the range of target parameters, errors caused by the nonuniqueness of inversion results are avoided. Through simulation and field experiments, we verify the performance of this method. The experimental results show that compared with the traditional inversion algorithms, the inversion accuracy is higher and the inversion speed is three to four orders of magnitude faster.