Introducing machine learning methods to Geothermal Play Fairway Analysis in the Great Basin region of Nevada

Abstract:

We have embarked upon a project to integrate machine learning (ML) methods with the Nevada play fairway analysis to develop an algorithmic approach for evaluating geothermal resource potential in the Nevada Great Basin region and identifying undiscovered blind geothermal systems. In support of this effort, this study focuses on developing training sites and input data that are compatible with ML techniques by synthesizing geologic and statistical constraints. We are also introducing the convolutional neural network (CNN), a highly scalable data driven approach, to automatically predict favorability and identify signatures for detecting blind geothermal systems.

Proper application of ML techniques requires a reasonable number of training sites. Initially, an inventory of negative training sites is under development to balance out potential positive sites representing ~85 active geothermal systems. For this study, we are integrating detailed datasets such as discrete high-resolution gravity and magnetic maps that make it possible to resolve detailed structural constraints critical for characterizing and analyzing the variance of favorability at both negative and positive training sites. We leverage the CNN to parameterize complex relationships between data and label pairs and learn a compact representation of constraints for discriminating background signatures from favorable zones of critically stressed faults that control deep circulating hydrothermal fluids. To optimize a CNN, the model is first trained on a simplified exploration task of identifying collocated structures to distinguish positive and negative sites. Once a suitable detection accuracy is achieved, we then plan to evaluate the location performance of the model to produce probabilistic location maps to help resolve local-scale favorability at various labeled and unlabeled basins and/or prospects.