FPGA-accelerated deep neural network for real-time inversion of geosteering data

We implement an efficient deep learning-based inversion method using the Field Programmable Gate Array (FPGA) to support real-time geosteering. FPGA is a feasible hardware choice for the harsh downhole environ-ment. Traditionally, due to the limitation of downhole computing resources, sensing data are transmitted from the downhole to the surface, and geosteering inversion is performed on the surface. Using a deep neural network (DNN) as the inverse operator, a pre-trained DNN can be implemented on the FPGA. In this paper, we propose a DNN-based inversion method for the geosteering problem and conduct a comprehensive analysis to accelerate the FPGA implementation. The experimental results show that our proposed approach is computationally efficient with high accuracy and low power consumption. Compared to the corresponding CPU (GPU) implementation, our method is 7x (1.4x) more computationally efficient and 82x (1.3x) more energy efficient.