Physics-guided deep reinforcement learning framework for reliability-based design optimization of complex mechanical systems

Reliability-based design optimization (RBDO) is essential for the safe and economical operation of complex mechanical systems, whose time-dependent performance must continuously satisfy safety requirements under multi-source stochasticity and multiple progressive damage. However, conventional RBDO methods face an inherent tension between computational feasibility and search capability, and are typically confined to a single design objective. To address this, a physics-guided deep reinforcement learning (DRL) framework for RBDO is proposed. First, the coupled degradation mechanisms are mapped to the time evolution of the key performance indicators, recasting the design task as a Markov decision process over a bounded design domain. Second, a composite reward—integrating an objective-improvement term, a physics-guided degradation-trajectory shaping term, and a constraint-violation penalty term—is embedded into the policy update, enabling high-dimensional optimization without a limit-state gradient, an inner reliability analysis, or a surrogate model. Finally, since the objective is carried by the reward rather than the problem structure, re-parameterizing it switches the same workflow between reliability and lifetime maximization. The framework is validated on three cases of increasing degradation complexity—a high-dimensional benchmark, a wing-spar–rib system, and an aircraft drag chute lock mechanism—and instantiated with four DRL agents, providing a unified, gradient-free solution approach across multiple reliability-oriented design tasks.