Predicting Magnetization Directions Using Convolutional Neural Networks

Proper interpretation of magnetic data requires an accurate knowledge of total magnetization directions of the source bodies in an area of study. I examined the use of machine learning, specifically Convolutional Neural Network (CNN), to automatically predict the magnetization direction of a magnetic source body, given a magnetic map. CNN has achieved great success in other applications such as computer vision, but has not been attempted in the realm of magnetics. I simulated magnetic data maps with varying magnetization directions from a cubic source body, all subject to the same inducing field. Two CNNs were trained separately, one for predicting magnetization inclinations and the other for predicting magnetization declinations. I also investigated various CNN architectures and determined the optimal architectures for predicting inclinations and declinations. The method works by generating many magnetic data maps with different magnetization directions as the training data for the CNNs. In order to generate these data maps, the user needs to interpret the source body parameters from the data. In this study, I also investigated how different source body parameters can affect the prediction of magnetization directions. My study shows that machine learning holds great promise for automatically predicting magnetization directions based on magnetic data maps. Two methods to increase the accuracy of the predictions are also explored in this study. The first method is to diversify the training data set. The second method is to use U-net, another type of CNN, to interpret the shape and lateral position needed to generate the training set. Both methods have proven well in improving the accuracy of the predictions and automating the process.