November 17, 2026
Journal Article
Deep Multitask Learning for Real-Time 3D Magnetic Data Inversion with Anomaly Segmentation
Abstract
We present a lightweight fully convolutional neural network (FCN) for 3D inversion of magnetic data, employing a multitask learning approach to simultaneously predict voxel-wise magnetic susceptibility and segment the subsurface into anomalous and non-anomalous cells. The proposed FCN framework is highly flexible and can be applied directly to various survey grid configurations and cubical mesh discretizations without modification. We train and evaluate our FCN models on a large synthetic dataset tailored for near-surface applications. On test data, our best-performing models recover over 95% of anomalous cells (within a 0.25m spatial tolerance) with more than 96% precision, while also producing accurate susceptibility estimates. The models remain robust under noisy conditions, maintaining high performance with up to 10% Gaussian noise added to the input magnetic data. We further evaluate our model using three field surveys conducted by partners at Exodigo Inc, demonstrating that our model correctly predicts the location and shape of buried utilities.Published: November 17, 2026