Magnetotelluric inversion using deep reinforcement learning with a smooth constraint

Inversion is one of the key steps in processing magnetotelluric sounding data and has been widely studied by scholars. The data-driven inversion approaches mainly include supervised inversion and semi-supervised inversion, but there is limited research on unsupervised inversion. ObjectiveDeep Q-network (DQN), a classical deep reinforcement learning algorithm, has recently been applied to one-dimensional magnetotelluric inversion problems as an unsupervised inversion approach. It has the advantages of not requiring a training dataset, being less dependent on the initial model, and being able to obtain the probability distribution of inversion results through multiple inversions. However, it suffers from the problem that the inversion results are not concentrated. MethodThis paper proposed a magnetotelluric inversion method based on deep reinforcement learning with a smooth constraint (SDQN). This method was based on the framework of reinforcement learning, treated the inversion problem as a Markov decision problem, and defined terms such as environment, reward, agent. Then, the model constraint term of regularized inversion was introduced into the reward function, guiding the agent to continuously adjust the resistivity parameters of the prediction model to obtain results that were more consistent with the model constraints. ResultsThe inversion results of the synthetic model showed that, compared with the DQN inversion and Occam inversion methods, the SDQN method produced more stable results when inverting observed data at different noise levels under the same number of iterations. The inversion results of the magnetotelluric measured data in the Tashikang mine area of Xizang were largely consistent with the Occam inversion results and aligned with the existing geological interpretations. ConclusionThe SDQN method has the advantages of more concentrated inversion results and stronger noise resistance to observed data, making it a new tool for solving the problem of magnetotelluric inversion.