Novel Machine Learning methods and tools for geothermal and geochemical problems

Machine learning (ML) methods are critical for improving our understanding of complex geoscience processes using available data. ML tools are also important for accelerating the data analyses and interpretations of large and complex datasets. ML can be applied to (1) extract hidden data features characterizing complex geoscience processes, (2) discover previously unrecognized interdependencies and mechanics, and (3) predict future states and conditions. Frequently, many geoscience datasets are large, uncertain, and contain measurement errors. The ML analyses can be substantially accelerated by providing physics (science) information in the ML process. This information can be provided in the form of constraints, mathematical expressions, or numerical models. They can be provided in the loss function utilized during ML optimization, or directly embedded in the ML methodology (for example, in deep neural networks).

Here, we demonstrate the application of physics-informed ML methods to geothermal and geochemical problems.

Major challenges associated with geothermal exploration and production are related to the complexity and uncertainty in the subsurface conditions. To mitigate these issues, we have developed and applied GeoThermalCloud.jl, a ML tool capable of mining public and proprietary datasets to estimate regional geothermal prospectivity, optimize field activities, design and site production wells, and more accurately predict geothermal energy production. GeoThermalCloud.jl is a part of our SmartTensors framework (http://smarttensors.com).

Geochemical processes involve complex interactions between fluid and solid constituents, critical for understanding contaminant fate, transport, and remediation. The development of reactive geochemical transport models predicting these processes from first principles is challenging and typically requires extensive model calibration (inversion) against observed site data. In addition, existing numerical geochemical models notoriously overestimate reaction rates in complex heterogeneous flow fields. We demonstrate how our ChemML tool can be applied for optimal remediation control by providing fast, robust, and defensible reduced-order geochemical models predicting contaminant transport based on the available site data.