Abstract

3D inversions, especially for potential field data that do not have the resolution of depth, are subject to severe non-uniqueness problems. While the non-uniqueness can be mathematically overcome by regularizations, a joint inversion of multiple geophysical datasets can significantly constrain the inversion model on a physical basis. Most conventional joint inversions are model-driven, through additional constraints appended to the objective functions as regularization terms. In this work, we explore the feasibility of joint inversions using a machine learning method by taking advantages of the recent boom in the artificial intelligence, particularly deep learning. In this study, we attempt to solve the joint inversion problem through the fully convolutional network (FCN) architecture based on the U-net workflow. First, a large number of randomly-shaped ellipsoid ore body models are generated, then the models are assigned appropriate density, susceptibility and electrical conductivity values and forward modelled to generate corresponding gravity, magnetic and audio-frequency magnetotelluric (AMT) data. Those random samples are used to training our neural network, whose input consists of 42 channels of 2D data arrays – one for gravity, one for magnetic and forty for AMT, and whose output is a discretized 3D model predicting the degree of mineralization in each cell. Our experiments show that deep learning-based joint inversions are feasible. Additional experiments were performed to verify whether data imbalance affects the training and performance of neural networks. We have found that the amount of data has little impact on the accuracy of the predictions.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call