Deep learning to estimate the basement depth by gravity data using a feedforward neural network

We have developed a deep-learning method based on the neural network of the feedforward type to estimate the depth to the basement from potential fields. The data used to train and test the network are related to the Bishop synthetic model. A trial-and-error approach was used to find the hyperparameters that have the best compromise between performance and computation time. The training was organized by associating the depth values of the basement to the data through a moving window, running along profiles in the north–south and east–west directions. In this way, we generated a set of approximately 296,980 examples. We verified the robustness of the trained net by carrying out a test related to another synthetic model, extracted from the Himalaya digital elevation model. The inherent ambiguity of the problem led us to test two hypotheses for the estimation of the basement depth, the first related to a priori information on the density contrast and the shallowest depth and the second assuming the knowledge of the depth at least at two points, but not that of the density contrast. In these cases, our data-driven approach yielded interesting results leading to estimate the maximum depth in the first case and the density contrast in the second one. We finally applied the method to the isostatic anomaly of the Yucca Flat sedimentary basin, Nevada. The results are consistent with previous interpretations of the area, which were based on gravity inversion methods.