Data-Driven Bicycle Design using Performance-Aware Deep Generative Models

Abstract

This treatise explores the application of Deep Generative Machine Learning Models to bicycle design and optimization. Deep Generative Models have been growing in popularity across the design community thanks to their ability to learn and mimic complex data distributions. This work addresses several key bottlenecks in the developing field, such as performance-aware generation, inverse design, and design validity. To support development of deep generative models, this treatise develops a foundation for data-driven design of bicycles, introducing three datasets: BIKED, BIKER, and FRAMED, considering holistic bicycle design, aerodynamic optimization, and structural optimization of bicycles respectively. It further proposes a set of tractable bicycle design tools, such as surrogate models to rapidly estimate performance of design candidates, analysis tools to guide the design process, and targeted design refinement tools using counterfactual explanations. This treatise finally proposes the first Deep Generative Model that actively optimizes for realism, performance, diversity, feasibility, and target satisfaction simultaneously. The proposed model achieves sweeping improvements over numerous evaluation criteria when compared to existing methods and establishes state-of-the-art performance on the bicycle design problem.