A deep learning approach to design a borehole instrument for geosteering

Abstract

Deep Neural Network (DNN)-based methods are suitable for the rapid inversion of borehole resistivity measurements. They approximate the forward and the inverse problem offline during the training phase and they only require a fraction of a second for the online evaluation (aka prediction). Herein, we propose a DNN-based iterative algorithm to design a borehole instrument such that the inverse solution is unique for a given earth parametrization. We select a large set of electromagnetic measurement systems routinely employed in logging operations, and our proposed DNN algorithm selects a subset of measurements that are suitable for inversion purposes. Numerical results with synthetic data confirm that this approach can provide valuable insight when designing borehole logging instruments.