A Machine Learning Framework for Estimating Peak Temperatures in Thermally Altered Rock Masses

Abstract Elevated temperatures represent a major hazard for rock masses in engineering structures, most notably during tunnel or building fires. Reliable assessment of thermal damage requires accurate determination of the maximum temperature reached. In this study, we present and evaluate a machine-learning approach for estimating both surface and sub-surface temperatures in thermally altered rock. Lower Carboniferous greywacke, selected as a model rock due to its complex response to heating, was subjected to controlled laboratory experiments. Input parameters included colour (in CIELAB colour space), bulk density, Shore hardness, and splitting tensile strength. A suite of machine-learning models was trained and compared with standard linear regressions. Our results show that machine-learning methods achieve up to four times higher accuracy (~ 10 °C versus ~ 40 °C) compared to previously used linear regressions, and remain effective even for rocks with heterogeneous composition. While colour is a reliable indicator for surface alterations, it is unsuitable for subsurface applications in greywacke; bulk density and mechanical properties provide more robust predictors. Geological variability of the greywacke was found to reduce prediction accuracy, but this can be mitigated through density-based corrections. In addition to methodological advances, the experiments highlight significant thermal instability of greywacke at temperatures above 1100 °C, raising concerns for its use in tunnel construction and fire safety.