Deep learning–based three-dimensional terrestrial temperature modeling throughout Japan incorporating multiple crustal properties and spatial correlation with an application to critical point distribution

Abstract

One of the most fundamental and essential issues for both ordinary and supercritical geothermal power generation is to specify hot zones with sufficient temperature and volume via the three-dimensionally estimated terrestrial temperature distribution from the surface to great depths over a wide area using well-temperature logging data. However, the amounts and locations of such data are generally limited, and target regions and depth ranges for temperature-at-depth predictions are generally narrow, and the targeted upper range of the temperature prediction is mostly lower than 250 °C. To overcome these problems, this study aims to enable an extensive temperature estimation by applying deep neural network (DNN) and neural kriging (NK) for targeting all of Japan. To supplement the temperature data, multiple crustal properties from geophysical and geochemical data, such as the Curie point depth, water quality of hot springs, active volcano distribution, and surface geology, are incorporated into the DNN and NK. The interpolation and extrapolation accuracies of the temperature logging data are evaluated using the holdout method, and the superiority of NK, especially for data extrapolation, is confirmed. Features of the NK temperature distribution that enable estimation of the temperature down to great depths with extremely sparse temperature logging data are determined and characterized. Using a three-dimensional temperature model, the distribution of the critical point of water is delineated throughout Japan and the expected resource densities are calculated under the condition of power generation using a steam flash system over 30 years. The results specify promising areas of supercritical geothermal resources, primarily located around typical active volcanoes, and large production power from a relatively shallow depth range over the long term.