Artificial intelligence-optimized shield parameters for soft ground tunneling in urban environment: A case study of Bangkok MRT Blue Line

This paper presents a study on multi-objective optimization (MOO) of shield operational parameters (SOPs) for soft ground tunneling using a tunnel boring machine (TBM) in an urban environment, focusing on the case study of the MRT Blue Line in Bangkok. The investigation aims to determine the optimal combination of SOPs, consisting of face pressure ($F_{\text{p}}$), thrust force ($T_{\text{f}}$), grout pressure ($G_{\text{p}}$), and percent grout filling ($G_{\text{f}}$), along with relevant environmental factors, including tunnel depth ($T_{\text{d}}$), inverted groundwater level ($W_{\text{i}}$), and type of surrounding soil ($T_{\text{s}}$). The primary objective is to enhance the penetration rate ($P_{\text{avg}}$, in terms of average value), as cost consideration, while mitigating ground surface settlement ($S$), as safety (serviceability) consideration. Using long short-term memory (LSTM) neural networks as predictive models, the results yield coefficient of determination (R²) values of 0.81 and 0.96, root mean square error (RMSE) values of 5.91 mm/min and 3.09 mm, and average bias factor values of 0.99 and 0.88 for the $P$ and $S$ predictive models, respectively, based on validation datasets. This integrated framework, which combines the non-dominated sorting genetic algorithm (NSGA-II) with LSTM neural networks, is applied to MOO to identify the optimal SOPs, while accounting for their influence on $S$ variation as a time-series over 11 timesteps, as considered in this study. For simplification and practical field implementation, the same set of SOP values is applied across all 11 timesteps during the optimization process. Using the proposed optimization framework, the optimal results demonstrate improvements in $P_{\text{avg}}$, increasing by up to 109.8% (from 13.99 to 29.35 mm) and in $S$, reducing up to 79.6% (from 34.55 to 7.06 mm) when MOO is conducted as a time series using the simplified method. This finding provides a valuable approach to effectively address the sequential uncertainties of relevant factors in soft ground tunneling for similar projects.