A self-learning framework for adaptive free-form pipe routing in aeroengines

Abstract

Pipe routing in conventional aeroengine design remains a complex and time-consuming manual task. Although considerable research has been devoted to automating the routing of constant-curvature pipes, the growing demand for free-form pipes introduces new challenges. The dynamic nature of design environments and diverse layout rules further increase the optimization complexity. To tackle these challenges, this study proposes a self-learning pipe routing (SLPR) framework for the automated design of free-form pipes. The framework represents pipe paths using cubic non-uniform B-spline curves to flexibly explore complex spatial domains. A discrete potential energy lookup map is proposed to encode the diverse layout rules. This map enables continuous path exploration with instantaneous rule evaluation through an efficient indexing function. For path optimization, SLPR employs the learning-based proximal policy optimization algorithm. Through iterative interaction with the environment, the SLPR agent refines its routing policy and accumulates reusable knowledge. This knowledge enables the agent to adapt to design modifications through fine-tuning, avoiding the computational overhead of retraining from scratch. Comparative tests reveal that SLPR effectively mitigates redundant segments existing in constant-curvature routing. In static environments, SLPR outperforms diverse algorithmic baselines in terms of path length, rule compliance, path complexity, and computational efficiency. In dynamic environments, the fine-tuned SLPR achieves faster convergence to superior solutions compared to retrained models. These results confirm that SLPR provides an efficient, intelligent, and robust solution for free-form pipe routing, with strong potential for real-world engineering deployment.