Reinforcement learning-based topology optimization for generative designed lightweight structures.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Computational Design

Background:

• Traditional generative design methods often struggle with complex constraints and manufacturability.
• Optimizing material layout for lightweight structures requires balancing performance and production feasibility.

Purpose of the Study:

• To develop an AI-driven generative design framework for lightweight, manufacturable mechanical structures.
• To integrate topology optimization with deep reinforcement learning for efficient material layout design.
• To ensure structural reliability and manufacturability through physics-informed learning and advanced processing.

Main Methods:

• Utilized deep reinforcement learning, specifically Proximal Policy Optimization (PPO), for topology optimization.
• Incorporated Finite Element Analysis (FEA) for physics-informed training and constraint adherence (Von Mises stress ≤ 300 MPa, displacement ≤ 0.5 mm).
• Employed Signed Distance Field (SDF) smoothing for manufacturability and generated STL files for 3D printing.

Main Results:

• Achieved up to 40% weight reduction compared to conventional methods like SIMP and level-set techniques.
• Demonstrated superior performance in maintaining structural compliance under defined engineering constraints.
• Successfully validated through a case study on a lightweight wheel hub and the Topology Optimization Dataset (ToD).

Conclusions:

• The AI-driven framework effectively generates lightweight and manufacturable mechanical designs.
• The integration of PPO, FEA, and SDF smoothing enhances design optimization and real-world applicability.
• The method offers a rapid design-to-prototype transition suitable for diverse engineering applications.