Machine-learning inversion via adaptive learning and statistical sampling: Application to airborne micro-TEM for seismic sand corrections

Abstract

We develop a novel physics-adaptive machine-learning (ML) inversion scheme showing optimal generalization capabilities for field data applications. We apply the physics-driven deep-learning inversion to a massive helicopter-borne transient electromagnetic (TEM) field data set. The objective is the accurate modeling of the near surface for enhancing the exploration of low-relief structures in a sand covered desertic area. Enhanced generalization of a neural network (NN) or other ML techniques is obtained from automatic physics-based adaptive training using images of real-world data. Data reduction schemes, based on statistical sampling techniques, enable the use of small and fully informative data sets for accelerated training. Adaptive learning is implemented via an iterative physics-driven data augmentation strategy. A deterministic inversion is regularized by a penalty term built from the difference between the inverted models and the models predicted by ML. The resulting inverted models and calculated responses are used for augmenting the ML training base. The automated procedure converges rapidly, producing a trained network model, which captures the general background data statistics as well as the characteristics of field-specific geophysical data as embedded in the augmented samples. The high-resolution ML inversions with acceptable levels of data misfit are obtained for large volumes of the field geophysical data. We test several different learning models, such as artificial NNs, convolutional NN (U-Net), and the Gaussian process, which provide stable and comparable results with network-specific characteristics. The developed scheme is successfully applied to micro-TEM data providing accurate near-surface seismic corrections for the exploration of low-relief structures.