Petrological controls on the engineering properties of carbonate aggregates through a machine learning approach.

Rock aggregates have been extensively exploited in the construction sector, and the associated engineering features play a critical role in their application. The main aim of this research is to assess the impact of petrographic characteristics on the engineering properties of carbonate rocks. A total of 45 carbonate rock samples from different geological formations within the Salt Range (Western Himalayan Ranges, Pakistan) were subjected to comprehensive petrographic analyses and standard aggregate quality control tests. The engineering characteristics encompassed Los Angeles abrasion value, aggregate crushing value, aggregate impact value, specific gravity, water absorption, and unconfined compressive strength, whereas petrographic examination of thin sections quantified the mineralogical composition. Statistical methods and machine learning models have been applied to elucidate the relationships between the petrographic and engineering features of the aggregates and establish potential predictive capability. The analysis identified clay, calcite, feldspar, and dolomite as the primary determinants for the engineering behavior of carbonate aggregates. Although multiple regression analyses produced R² values exceeding 0.84, the multiple regression equations did not provide substantial insights into the impact of all petrographic parameters on engineering properties. To enhance predictive accuracy, advanced machine learning models, including Random Forest, Gradient Boosting, Multi-Layer Perceptron, and Categorical Boosting, were applied. Among these, the Gradient Boosting model demonstrated superior predictive capability, overcoming both traditional regression methods and other machine learning algorithms as validated through the Taylor diagram and ranking system (i.e., r = 0.998, R² = 997, Root mean square error = 0.075, Variance Accounted For = 99.50%, Mean Absolute Percentage Error = 0.385%, Alpha 20 Index = 100, and performance index = 0.975). These results highlight the ability of machine learning techniques to provide a more effective and reliable prediction of aggregate engineering properties based on petrographic data. This approach offers significant advantages in the preliminary assessment of aggregate suitability, contributing to more efficient resource allocation in construction projects.