Preliminary Performance Assessment of a Customized GPT Model for Geomechanical Classification of Rock Masses Using Images and Input Data

Abstract

Geomechanical characterization of rock masses is essential for the stability of mining and civil works; however, it is still mostly performed manually, making the process time-consuming and highly dependent on human judgment. This study presents a preliminary evaluation of the performance of a customized GPT model for geomechanical classification using images and input data, comparing its results with field classifications obtained using the Bieniawski RMR system. The research followed a quantitative approach with an applied level, a non-experimental and cross-sectional design, and an exploratory and descriptive scope, analyzing 30 observation points from rock outcrops in the Cajamarca region (Peru). The model evaluated each image in three independent runs and was configured using the technical criteria of the RMR system. The results showed an average accuracy of 83.4%, a simple Kappa of 0.76, and a weighted quadratic Kappa of 0.88, corresponding to substantial to almost perfect agreement, with an overall inter-run reliability of 75.6%. Overall, the customized GPT model demonstrated solid preliminary performance and results consistent with human classification, suggesting potential for future application in the geomechanical characterization of rock masses.