Learning hyperparameter predictors for similarity-based multidisciplinary topology optimization.

Area of Science:

• Engineering
• Computational Science
• Materials Science

Background:

• Topology optimization (TO) is crucial for industrial design, offering optimal material distribution based on design space and loads.
• Current TO methods struggle with user control over final designs, especially in multidisciplinary tasks with conflicting criteria.
• Existing similarity-based TO methods require extensive hyperparameter sampling, increasing computational expense.

Purpose of the Study:

• To develop a novel end-to-end approach for similarity-based topology optimization.
• To integrate a machine learning model for predicting hyperparameters, reducing computational cost.
• To provide engineers with greater control over TO designs, ensuring similarity to reference structures.

Main Methods:

• Proposed a novel end-to-end similarity-based topology optimization (TO) methodology.
• Integrated a machine learning model to predict hyperparameters for TO methods.
• Generated training data from an academic linear elastic problem for model generalization.

Main Results:

• The machine learning model successfully predicts hyperparameters, enabling similarity-based TO at minimal computational cost.
• The approach demonstrates generalization across linear elastic, nonlinear dynamic crash, and industrial-scale TO problems.
• Successfully applied to a real-world car hood frame design problem, validating industrial applicability.

Conclusions:

• The proposed machine learning integrated approach enhances user control in topology optimization.
• This method significantly reduces the computational burden associated with achieving desired design similarities.
• The approach is effective for real-world industrial multidisciplinary design problems, such as automotive components.