Reinforcement learning-driven dynamic optimization strategy for parametric design of 3D models.

Area of Science:

• Computer-Aided Design (CAD)
• Computational Geometry
• Artificial Intelligence (AI)

Background:

• Parametric design enables precise manipulation of complex 3D forms in architecture, fabrication, and product design.
• Optimizing large, coupled parameter spaces in 3D modeling presents significant computational challenges.
• Existing methods struggle with the efficiency and scalability required for complex design exploration.

Purpose of the Study:

• To introduce a novel Hierarchical Reinforcement Learning based Dynamic Optimization Strategy (HRL-DOS) for 3D parametric design.
• To address the computational challenges of exploring and optimizing large parameter spaces in complex 3D modeling.
• To enhance the efficiency and adaptability of automated parametric design processes.

Main Methods:

• Decomposition of the parametric design process into a series of multi-level subproblems using HRL-DOS.
• Implementation of a high-level policy for global design direction and a low-level policy for parameter adaptation.
• Integration of multiple performance criteria, including structural stability, geometric efficiency, and fabrication constraints.

Main Results:

• HRL-DOS demonstrated a 27% improvement in convergence speed compared to heuristic or gradient-based methods.
• An 18% improvement in the quality of the 3D models was observed using the HRL-DOS approach.
• The hierarchical strategy showed enhanced learning efficiency and computational scalability in complex design environments.

Conclusions:

• HRL-DOS offers a new, adaptive, and efficient approach to automating parametric design tasks in 3D modeling.
• The strategy's hierarchical decomposition effectively manages complex parameter spaces and multiple performance criteria.
• This method holds potential for applications in architectural form-finding, generative product design, and intelligent CAD systems.