Improving the Interpretation of Magnetic Tensor Data Using Deep Learning

The accurate interpretation of magnetic tensor data can be difficult to perform without a strong knowledge of local geology and experience in reading magnetic data. I examined ways in which machine learning techniques can be applied to magnetic tensor data to automatically locate possible kimberlite targets and a method to sharpen smoothness based inversion models to provide a clearer image of the subsurface. While machine learning networks like the U-Net have shown success in other fields for image processing, these methods have not been used extensively in geophysics for magnetic interpretation. I trained the U-Net to predict kimberlite pipe locations by forward modelling magnetic susceptibility models to corresponding magnetic tensor data. I examined the use of different neural network architectures and methods for calculating loss to determine the effect on prediction accuracy. The U-Net was adapted in order to sharpen inversion models by changing from a two dimensional layer architecture to three dimensions. To train this second U-Net I first created a smoothing function which closely matches the effects of a smoothness based inversion and then applied this smoothing function to three dimensional magnetic susceptibility models. I also compared the effectiveness of using a general smoothing function against a smoothing function specifically designed to match a smoothness inversion. This study shows that the use of the U-Net architecture in the field of magnetics shows great promise in the areas of automatically detecting targets and sharpening inverted models.