Cerchar abrasiveness index prediction based on rock properties leveraging hybrid soft computing techniques

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence
• Rock Mechanics

Background:

• The Cerchar Abrasiveness Index (CAI) is critical for assessing rock abrasiveness in tunneling and mechanized excavations.
• Accurate CAI prediction aids in selecting appropriate TBM cutters and optimizing excavation efficiency.
• Traditional methods for CAI determination can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop and validate an accurate and efficient AI-driven methodology for predicting the Cerchar Abrasiveness Index (CAI).
• To compare the performance of various machine learning algorithms optimized with metaheuristic techniques for CAI prediction.
• To assess the engineering applicability of the proposed models using real-world tunneling project data.

Main Methods:

• Utilized a dataset of 163 rock samples with diverse geological origins (igneous, sedimentary, metamorphic).
• Employed base algorithms (XGBoost, LightGBM, Random Forest) enhanced by metaheuristic optimizers (AOA, RSO, HHO).
• Input parameters included Brazilian tensile strength (BTS), uniaxial compressive strength (UCS), equivalent quartz content (EQC), and brittleness index (BI).
• Model performance was evaluated using metrics like R², RMSE, MAE, and VAF, with data split into 80% training and 20% testing sets.
• External validation was performed against 17 international hard rock TBM projects.

Main Results:

• The AOA-optimized models demonstrated superior performance, with AOA-LightGBM achieving R² = 0.952 and AOA-XGBoost reaching R² = 0.951 on the test set.
• External validation showed AOA-XGBoost had a high correlation (0.8308) with field measurements from tunneling projects.
• Feature importance analysis revealed EQC as key for XGBoost, while UCS was most influential for LightGBM and Random Forest models.
• The developed models proved more accurate and efficient than traditional experimental methods.

Conclusions:

• The proposed AI-driven approach, particularly AOA-optimized models, offers a significant advancement in CAI prediction methodology.
• The models exhibit high accuracy, efficiency, and broad applicability across various rock types.
• Validated against real-world data, the methodology demonstrates strong engineering potential for mechanized tunneling projects.