The design of real-world systems—whether in engineering, architecture, software, industrial, financial, or social domains—is often a challenging and dynamic process marked by remarkable successes and occasional failures. Historically, it has been widely believed that only human experts possess the ability to comprehend large and complex problems, conceptualize innovative solutions, and map functional design requirements to real world entities. This raises an intriguing question. When will machines be capable of independently thinking and designing complex systems, referred to as a Thinking Design Machine (TDM)?
Some might argue that we are already on the path to TDM with the advent of generative design tools. These tools enable designers and engineers to explore countless possibilities for product parameters while optimizing for performance, aesthetics, and economics based on narrowly specified constraints and algorithms set by human designers. Despite their impressive capabilities, current generative design tools are limited to parameter variations, such as topological optimization, at very high speed, since they lack the ability to understand problems, conceptualize new ideas, or create entirely novel functions.
The recent success of AlphaFold in predicting over 200 million protein structures demonstrates the potential of advanced machine learning technologies to solve complex problems in proteomics, when domain specific principles and algorithms can be provided. Inspired by such breakthroughs, we aim to explore pathways for enabling TDM by uncovering functional design principles and algorithms alongside state-of-the-art machine learning techniques.
