Hybrid Surrogate - Physics - AI Optimization of EPB-TBM Performance: A Data-Driven Application from the Jakarta Metro Project

Maharani Rindu Widara (Mining Engineering, Politeknik Batulicin, South Kalimantan, 72271, Indonesia)
Chairul Salam M. (Istanbul Technical University, Maslak, Istanbul, 34469, Turkey)

Jurnal Teknik Industri: Jurnal Keilmuan dan Aplikasi Teknik Industri
Vol. 28 No. 2 (2026): December 2026 (Pre-Print)

The increasing complexity and geological unpredictability associated with mechanised tunnelling demand optimisation frameworks that go beyond static, offline parameter adjustments and embrace adaptive and physically coherent decision-making processes. Traditional metaheuristic algorithms, such as genetic algorithms (GA), particle swarm optimisation (PSO), and simulated annealing (SA), have been widely applied to optimise tunnel boring machine (TBM) performance. However, they are inherently limited by their lack of real-time adaptability and restricted physical interpretability. This research presents a hybrid surrogate–physics–AI optimization framework (H-SGP-BO–PINN–DRL) that redefines TBM optimization as a physics-constrained adaptive control challenge, shifting away from a static optimization model. The framework integrates a surrogate-assisted hybrid GA–PSO combined with Bayesian optimisation for a global search that considers uncertainty, a physics-informed neural network (PINN) that incorporates essential mechanical principles to ensure geotechnical feasibility, and a deep reinforcement learning (DRL) controller that enables real-time adjustments within physically feasible operational limits. The framework was validated using field data from EPB-TBM projects in the Jakarta MRT and Istanbul Metro systems, as well as synthetic datasets generated using FEM. The results show that the proposed method reduces the specific energy consumption (SEC) and improves the penetration rate (PR), while also demonstrating superior convergence stability and robustness compared to traditional metaheuristic and surrogate-based methods. Monte Carlo bootstrapping and Sobol global sensitivity analysis further identify thrust force and cutterhead rotation speed as the main factors affecting energy-performance variability. By integrating surrogate-assisted optimization, physics-informed learning, and reinforcement-based adaptability, this study introduces a new paradigm for intelligent TBM operation, enabling interpretable, energy-efficient, and deployable decision support for future smart and autonomous tunneling systems.