Deep learning-powered rock mass classification: Predicting RMR from Q-system parameters with high accuracy

Reliable stability assessment requires an objective and precise assessment of the rock mass quality classification. A deep learning model is developed to create a tool that can provide a rapid and precise assessment of the quality of rock masses. While there are empirical equations to determine RMR values from Q parameters, this study provides an advanced highly accurate deep learning approach to determine RMR values from Q parameters. This serves to reduce the amount of fieldwork related to collecting the rockmass data needed to independently assess rockmass quality using the RMR system and the Q system separately. The RMR values, like Q values, were first determined independently in the field. The deep learning approach was later used to predict the field-determined RMR values from the field-determined Q parameters. This means that each practical field measurement point had an RMR, and a Q value independently determined for it before the deep learning approach was applied. The six rockmass parameters of the Q system (RQD, J n , J r , J a , J w , SRF) are used as input in this model while the RMR is used as the output variable. In this study, the dataset contains 356 samples, 70%, 15% and 15% of the entire sample data are used to train, test, and validate the model, respectively. The predictive performance of the models was evaluated and compared using metrics such as R 2 , MAE, and RMSE among many others. The overall R 2 values for the ANN, FDA-ANN and SCA-ANN are 0.9951, 0.996 and 0.9955 respectively. The MAE values are 0.099, 0.096 and 0.085 for ANN, FDA-ANN and SCA-ANN respectively. The FDA-ANN model has a higher accuracy than other techniques, such as the ANN and SCA-ANN. The error values obtained for each of the models are very close to their expected value of 0 while their obtained R 2 and VAF are also much closer to the targeted value of 1 and 100% respectively. The PI is also close to the expected value of 2. Hence, the three proposed models can be confidently used in predicting RMR values using Q parameters obtained from field investigations without the need to independently determine RMR from the traditional RMR field parameters. The study used the Chord diagram to display the rank of the performance indicators and the sensitivity analysis using the Cosine Amplitude methods (CAM). It shows that the RQD parameter has the highest CAM value followed by J w and then J n for all three models. The results offered here provide insight for engineers and academics who are interested in analysing rock mass classification criteria or conducting field investigations. • Developed a deep learning-based tool to predict RMR values from Q-system parameters, reducing the need for extensive fieldwork. • Compared ANN, FDA-ANN, and SCA-ANN models, with FDA-ANN achieving the highest accuracy (R 2 = 0.996, MAE = 0.096). • Used Chord diagrams for performance ranking and Cosine Amplitude Methods (CAM) for sensitivity analysis, identifying RQD, Jw, and Jn as the most influential parameters. • Demonstrated that AI-driven models outperform traditional empirical equations in RMR estimation with high precision. • Provides a reliable and efficient alternative for geotechnical engineers and researchers in rock mass classification.