Development of an integrated optimization method for shape generation and material selection using machine learning

Complex mechanical products are composed of thousands to millions of parts, making it essential to utilize numerical calculations such as finite element analysis to improve performance and reduce costs. Recently, advancements in computer performance have promoted the optimization of component shapes, however, a trade-off between the flexibility of shape generation and adherence to design constraints has emerged as a challenge. For example, topology optimization using density methods can yield diverse shapes but often ignores manufacturing constraints. Thus, current methods struggle to balance the exploration of diverse shapes with compliance to manufacturing constraints. Furthermore, the relationship between shape and material selection is also crucial in design. When using multiple materials, the properties of the materials will influence shape design, however, optimization methods like density methods fix material properties during calculations, making it difficult to consider both shape and material concurrently in the optimization process. As a result, designers often rely on experience to select materials and shapes. This study proposes a method that simultaneously allows for high flexibility in shape modifications aligned with the engineer's intentions while optimizing materials and composite structures. In this method, the displacement of node coordinates in finite element analysis is defined as a morphing vector and generates diverse shapes by varying the magnifications of these vectors. By employing Bayesian optimization to minimize stress and maximize flexural rigidity, material selection is also optimized concurrently with shape variation. The optimization calculation is substituted for a machine learning surrogate model to estimate stress in order to keep computational cost low.