We are applying machine learning techniques, including artificial neural networks and training site augmentation, to mitigate key challenges in the Nevada geothermal play fairway analysis. The ultimate goal is to develop an algorithmic approach to identify new geothermal systems in the Great Basin region. The specific objectives of this study are to 1) develop training sites and input data that are compatible with machine learning techniques by synthesizing geologic and statistical constraints; and 2) introduce a convolutional neural network model to automatically predict favorability at sites within the play fairway study area and identify signatures to detect blind geothermal systems. Initially, an inventory of negative training sites is being developed to balance out potential positive sites of ~85 active geothermal systems. Future work will review datasets associated with temperature and permeability to better define the variance of favorability at both negative and positive sites and explore the influence of machine learning applications, in particular the convolutional neural network model, to parameterize complex relationships between data and label pairs. To optimize a convolutional neural network, the model is trained on a simplified exploration task of identifying collocated structures from synthetic data as well as geologic and geophysical data; this model may help identify the location of critically stressed faults prone to slip and dilation that control deep circulating hydrothermal fluids. Results of this study include 1) the development of training sites compatible with machine learning techniques; 2) a review of data selection and machine learning augmentation techniques; 3) prediction statistics with different numbers of convolutional neural network layers to determine the optimal model architecture that achieves the highest testing accuracy; and 4) detailed local favorability maps derived from an optimized convolutional neural network model.
Machine Learning in Geothermal Development