Abstract
This study presents an AI-driven generative design framework for creating lightweight, manufacturable mechanical structures. It integrates topology optimization with deep reinforcement learning, specifically the Proximal Policy Optimization (PPO) algorithm, to learn optimal material layouts within a defined design space. The model adheres to strict engineering constraints, including Von Mises stress (≤ 300 MPa) and displacement (≤ 0.5 mm), ensuring structural reliability. Physics-informed learning is enabled through Finite Element Analysis (FEA), enhancing the model's decision-making during training. To improve manufacturability, the framework applies Signed Distance Field (SDF) smoothing and generates STL files suitable for direct 3D printing. Tested on the Topology Optimization Dataset (ToD), the method outperforms conventional approaches like SIMP and level-set techniques, achieving up to 40 % weight reduction while maintaining compliance. A practical case study involving a lightweight wheel hub further validates its real-world applicability. Comprehensive evaluations, including ablation studies and inference-time analysis, demonstrate the method's adaptability, constraint satisfaction, and rapid design-to-prototype transition across engineering domains. Methodology summary includes:•AI-based generative design with PPO under mechanical constraints.•Physics-informed training with FEA and SDF-based STL output.•Evaluated on ToD and validated through a wheel hub case study.