Vision transformer networks for airborne IP removal

We propose a machine-learning model that can predict the time-domain Electromagnetic (EM) data from airborne acquisition in the presence of Induced Polarization (IP) effects. The model uses a Transformer network to map our time-domain EM+IP input to a pure EM output, which may or may not contain IP effects. Transformers’ advantage over other networks is the Attention Mechanism (Vaswani et al., 2017), which, in computer vision, relates a single pixel with every other pixel in an image. It is not limited to only nearby-neighboring pixels. We can train a network to identify the EM response in the time-domain EM data that contains IP signals using spatial and temporal information. The network processes input data like an image, where stations and windowed time data make up the pixels of the input. We train on synthetic data and show that we can recover conductivity models similar to the EM-only signal.