Enhanced generalization of a deep learning framework for 1-D magnetotelluric inversion

Abstract

Widespread and successful application of deep learning‑based methods to solve inverse problems in Geophysics particularly - in magnetotellurics (MT) – remains limited largely due to poor generalization of trained models to unseen and noisy field data. In order to contribute to addressing this issue, we present a robust one‑dimensional (1‑D) MT inversion framework in which the deep learning model has increased capacity to generalize. First, by employing a randomized conductivity model construction scheme, constrained by geophysical resolution limitations and geologically realistic scenarios, a comprehensive and diverse training dataset representing a wide variety of subsurface conductivity models is generated. To further ensure a representative sampling of the model space, Gaussian noise is added to part of the generated data. Second, a clustering‑based data splitting strategy is introduced as an alternative to the random approach, in which each cluster proportionally contributes to the training, validation, and test subsets, to preserve consistent statistical distribution across them. Qualitative and quantitative evaluations confirm that the combined approach performs robustly and reliably when exposed to a broad range of synthetic data as well as a sample of real field MT data. Resistivity models constructed using a modified convolutional U‑Net architecture, trained with a combination of clean and noisy data generated in the stated manner, are compared to those derived from deterministic methods and show high levels of accuracy.