DESIGN SPACE EXPLORATION AND OPTIMIZATION OF MECHANICAL COMPONENTS USING MACHINE-LEARNING TECHNIQUES

Abstract

In today's highly competitive market, rapid product development with minimal resource utilization is crucial. This requires thoroughly evaluating all available design solutions to identify the most efficient options. While various techniques exist to explore design spaces and achieve optimal results, continuous research is dedicated to discovering more effective approaches that can be adapted to diverse design challenges. This thesis examines how supervised machine-learning techniques can be used to explore and optimize mechanical component design problems. Specifically, it focuses on three mechanical component design problems: pressure vessel design, helical coil spring design, and belt pulley drive design. Four machine learning classification models (support vector machine, random forest, gaussian naïve Bayes, and neural network) are tested. We prepared three different dataset sizes for both binary and multiclass classification using simulation-based design of experiments to investigate how dataset size affects model performance. We used the Latin hypercube sampling method to effectively sample points from the available design space. Additionally, hyperparameter tuning was performed to improve the performance of the evaluated models. Based on our findings, the random forest and support vector machine models outperform the others. Specifically, the random forest model excels in all three design problems for binary and multiclass classifications across various dataset sizes, even with default parameters. However, the support vector machine and neural network model can surpass the random forest's performance when hyperparameters are fine-tuned. On the other hand, the Gaussian naïve Bayes model exhibits the lowest accuracy in all three design problems. Interestingly, regardless of dataset size, there are no significant variations in the classifiers' performance for both binary and multiclass classifications. This suggests that the classifiers' effectiveness relies more on the dataset's representation of the original distribution than its size. This implies that reducing the sampling budget is possible using a small number of data points that accurately represent the design space. This study shows how machine learning classifiers efficiently solve mechanical component design issues, particularly in exploring design spaces and finding optimal values.