A framework for design optimization across multiple concepts.

Area of Science:

• Engineering Design
• Computational Science
• Machine Learning

Background:

• The initial concept design phase significantly impacts product life cycle costs (up to 70%).
• Standard optimization methods are ill-suited for the multi-concept design phase due to fixed-concept assumptions and computational demands.
• Existing methods struggle to efficiently search across diverse conceptual solutions with varying variable types.

Purpose of the Study:

• To develop an efficient framework for searching optimal solutions across multiple engineering design concepts.
• To address the gap in applying optimization methods during the early, multi-concept design phase.
• To minimize computational budget by progressively evaluating promising concepts and solutions.

Main Methods:

• A novel framework employing data-driven decisions for concept and solution evaluation.
• Integration of Tree-structured Parzen Estimator (TPE) with Gaussian Process (GP) and Random Forest (RF) regressors.
• Extension of GP and RF for multi-concept search and highlighting TPE benefits in design optimization.

Main Results:

• Demonstrated efficient search across multiple concepts with heterogeneous variable types (continuous, binary, discrete, categorical).
• Successful application in diverse case studies including cantilever beams, coronary stents, and lattice structures under limited computational budgets.
• Validation of the framework's ability to reduce computational cost and avoid premature convergence.

Conclusions:

• The developed framework effectively supports designers in the concept-based design phase.
• Machine learning advancements, particularly TPE, offer significant benefits for multi-concept design optimization.
• This approach fills a critical gap, enhancing efficiency and effectiveness in early-stage engineering design.