Deep Learning 3D Sparse Inversion of Gravity Data

Gravity prospecting is an important geophysical method for mineral resource exploration and investigating crustal structures. Based on the importance of this method, we propose a novel method that takes advantage of rock data, using a supervised deep fully convolutional neural network, that generates a sparse subsurface distribution from gravity data. During the data preparation phase, we used the random walk to synthesize diverse geological models, in which each model element has only two choices. During network training, we feed the geological model as labels and their corresponding forward modeling of gravity data as the input, after which the network parameters are learned using the Dice coefficient. During network testing, six general types of 3D models were developed, and corresponding gravity data was entered into a trained network to achieve the prediction results in less time. The statistical analysis of two evaluation metrics showed that our network was highly effective using our proposed dataset, wherein the recovered models were characterized by distinct boundaries. Furthermore, our approach was validated using real data obtained from the San Nicolas deposit in central Mexico. The files provided contain the training datasets used in the manuscript, as well as MATLAB scripts and Jupyter Notebooks to generate the training sets and predict 3D models