Predicting the CERCHAR Abrasiveness Index (CAI) from Geotechnical and Geophysical Rock Properties Using Machine Learning (ML) Regression Models

The CERCHAR Abrasiveness Index (CAI) is widely used to quantify rock abrasiveness, yet its contact-based scratch procedure is highly sensitive to operational factors such as scratch length, surface preparation, stylus hardness, confining pressure, scratch velocity, and observer variability. The requirement for a prepared, polished surface and styli of specific Rockwell hardness further complicates testing, making CAI difficult to standardize across laboratories and challenging to compile into consistent databases. This study evaluates whether CAI can be reliably predicted from geotechnical and geophysical properties typically obtained through in situ testing or routine core characterization. A dataset of 175 igneous, sedimentary, and metamorphic rocks was assembled from publicly available CAI databases and laboratory sources. A rigorous validation pipeline was implemented, including Leave-One-Out Cross-Validation, K-Fold and Repeated K-Fold validation, and hyperparameter tuning using Nested K-Fold across seven regression models. A four-predictor model using Schmidt hammer rebound hardness, Mohs hardness number, longitudinal wave velocity, and grain size is proposed. Across all tested seeds and rock types, the Random Forest model showed the most stable performance, achieving K-Fold R2 values of 0.94–0.98 with low RMSE and MAE. Because these predictors can be measured directly on outcrops, blocks, or tunnel faces, the approach enables a non-destructive, field-deployable CAI estimation without the need for prepared samples or specialized laboratory infrastructure.