Dynamic prediction models of rock quality designation in tunneling projects

Machine learning (ML) is becoming an appealing tool in various fields of civil engineering, such as tunneling. A very important issue in tunneling is to know the geological condition of the tunnel route before the construction. Various geological and geotechnical parameters can be considered according to data availability to define tunnels' ground conditions. The Rock Quality Designation (RQD) is one of the most important parameters that are very effective in tunnel geology. This article aims to maximize the prediction accuracy of the RQD parameter along a tunnel route through continuous updating techniques. For this purpose, four ML methods of K-nearest neighbor (KNN), Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Decision Tree (DT) were considered. All the RQD observations along the tunnel mute were considered as the models' inputs. For predicting the RQD status along the entire tunnel mute, the ML models use the regression technique. For checking the applicability of the models, the Hamru road tunnel in Iran was used. The models were updated twice to assess the update effect on the results achieved during the tunnel construction. In each prediction phase, all the prediction results were compared using different statistical evaluation criteria and the actual mode. Finally, the comparative tests' findings showed that predictions of the GPR model with R-2 = 0.8746/root mean square error (RMSE) = 3.5942101, R-2 = 0.9328/RMSE = 2.5580977, and R-2 = 0.9433/RMSE = 1.8016325 are generally well-suited to actual results for pre-update, first update, and second update phases, respectively. The updating procedure also leads to prediction models that are more accurate and less uncertain than the previous prediction stage.