Bayesian deep reinforcement learning for uncertainty quantification and adaptive support optimization in deep foundation pit engineering

Abstract

Abstract This study develops a novel framework integrating Bayesian inference with deep reinforcement learning for uncertainty quantification and adaptive support optimization in multi-physics coupled deep foundation pit systems. The methodology systematically incorporates prior knowledge and real-time monitoring observations through Markov Chain Monte Carlo updating to refine parameter estimates while employing deep reinforcement learning algorithms for autonomous support optimization. A comprehensive multi-physics coupled numerical model captures mechanical-hydraulic-thermal interdependencies with explicit coupling mechanisms and Shanghai-specific soil characterization. Validation through a representative Shanghai deep foundation pit project demonstrates superior performance with prediction accuracy (R2 = 0.91), reliability quantification (coverage probability = 96.8%), and practical improvements including 35% reduction in maximum wall displacement (from 45.8 to 29.7 mm), 42% decrease in surface settlement (from 28.5 to 16.5 mm), and 18% cost savings (¥2.3 million) compared to conventional deterministic approaches. The intelligent system achieved zero safety incidents, 12% construction duration reduction, and enhanced deformation control through real-time adaptive support adjustments including optimized prestressing forces and strategic anchor installations. The research contributes essential theoretical foundations and practical tools for advancing intelligent construction practices in complex urban geotechnical engineering applications.