The Bayesian Inversion Method With a Surrogate Modeling Based on Neural Network for GATEM Data

The ground-source airborne time-domain electromagnetic (GATEM) system is an efficient geophysical survey system for geological surveys and mineral surveys. The geological resistivity structure is obtained by inversion methods, and however, the deterministic inversion methods can only provide an optimal resistivity model. The Bayesian inversion method can provide the posterior probability distribution; however, it requires a large amount of calculation. In this article, to improve the efficiency, a surrogate modeling based on neural network (NN) is applied to replace forward simulation calculation in the Bayesian inversion method. The accuracy of the NN-based surrogate modeling is related to the training sample set. To obtain high-precision inversion results, the surrogate modeling based on NN will be update adaptively online. Above all, an initial NN-based surrogate modeling is trained on a sample set of prior information. The high-fidelity surrogate modeling based on NN is obtained through new sample sets of GATEM data that are generated to update the surrogate modeling, if the NN-based surrogate modeling is inaccurate when the Bayesian inversion method is running. An optimal solution model and the posterior probability distribution of model parameters for GATEM inversion results are calculated through the Bayesian method. The effectiveness of the Bayesian inversion method with a surrogate modeling based on NN is verified by the GATEM responses for typical geological models.