Data Driven Subsurface Workflow for Geothermal Resource Exploration: Applications of Machine Learning Methodology

Abstract As the drive towards an orderly energy transition intensifies globally, the urgency of finding and producing economical geothermal energy resources is becoming more important. Historically there is lack of dedicated exploration wells within potential geothermally active regions. Well logs, especially temperature logs, is a key element for geothermal subsurface interpretation workflows. In addition, acquisition of temperature well log is an expensive and time-consuming task for any drilling campaign. Hence, there is need for technological innovations to overcome the lack of available temperature well log data and simultaneously to ensure high confidence temperature profile prediction for prospective geothermal reservoirs. Subsurface temperature profile depends on many factors such as surface temperature, heat flow produced from mantle and crust, thermal properties of rocks, burial history, tectonics and faults, geochemical effects of circulating fluids, etc. Hence, incorporation of such diverse physical processes and relevant data sources are important to predict a high confidence regional temperature variations. Bottom-hole temperature (BHT) measurements are commonly used to map subsurface temperatures for geothermal gradient analysis. BHT data is primarily collected at different wells including shallow water wells, oil & gas wells and deep stratigraphic wells, where maximum temperature is usually reported at the depth of each drilled section. Based on this "point" data, one tends to leverage a simple thermal conductivity model coupled with stratigraphic knowledge to predict the subsurface temperature profile. This crucial information is further used to calculate the requirements of geothermal power plants construction as well as the drilling and completion design of subsequent wells in the field. In order to better predict such important parameter as subsurface temperature profile, machine learning (ML) coupled with geo-statistical methods hold promising potential. ML algorithms have the ability to learn from data and further generalize this learning to "unseen" data. In this way, ML helps to decipher complex relationships among input parameters, i.e. "features", which could be used to predict important reservoir parameters, i.e. the temperature profile at geothermal exploration targets. In this paper, we are proposing a supervised ML-based subsurface workflow for predicting the temperature profile within geothermally active areas. Our aim is to leverage existing data from O&G wells and near surface geological information to map non-linear relationships among physical parameters affecting geothermal gradient prediction. Within this scope, we would also like to demonstrate the usage of data-driven technologies to address the issues of determining missing well logs and how ML algorithms can be an enabler. We believe that our proposed data-driven workflow would enable automation within high grading of geothermal exploration targets on a regional scale. In this way, exploration teams could rapidly screen for geothermal anomalies, thus covering vast geothermal prospective areas within Saudi Arabia.