Rocks at High Pressure and Temperature: Bridging Index and Engineering Properties Using Physics-Guided Machine Learning

Abstract Heat recovery from superhot rock is a transformative, high-density energy source with strong baseload potential. However, in situ geomechanical characterization remains impractical, while conventional sampling and laboratory testing are intrusive, costly, and alter material properties. We compile a curated database of 1711 specimens spanning eight lithologies tested under high-pressure, high-temperature conditions to examine the use of index properties recoverable from drill cuttings—rock type, porosity, and mineralogy—to predict stiffness, strength, and brittle–ductile response at in situ stress and temperature. Physics-guided machine learning (PGML) embeds semi-empirical baseline equations as physical constraints while extracting high-dimensional patterns to derive explicit multiplicative formulations. This structure preserves phenomenological consistency beyond the calibration range and enables systematic refinement as new data emerge. The model predicts stiffness and strength with a normalized mean absolute error NMAE~0.3, and the brittle-to-ductile transition confining stress with 85% accuracy, while limiting the critical misclassification of ductile failure as brittle behavior to ~3%. Stiffness reflects formation history, with lithology, mineralogy, and porosity as key descriptors; in contrast, strength depends strongly on the current stress state. Some lithologies require additional microstructural descriptors beyond mineralogy, yet PGML predictions capture the governing microphysical processes encoded in the inputs. These results enable real-time drilling optimization and inform thermo-mechanical stimulation design.