Generative AI models like ChatGPT are rapidly becoming adept at replicating existing content, but MIT engineers believe there’s a significant gap when it comes to innovation, especially in the realm of engineering design. A new study from MIT claims that while these models are great mimics, they may lack the potential to truly innovate.
Lyle Regenwetter, a mechanical engineering graduate student at MIT, pointed out the inherent flaw of these models: their primary objective is to replicate a dataset. However, the essence of engineering is to innovate beyond what’s already available. The current focus on “statistical similarity” means that most designs produced by the models tend to closely resemble previous designs without necessarily improving performance.
In a detailed case study involving bicycle frame design, the researchers found that generative AI models often produced designs that, while resembling established frames, didn’t meet the necessary engineering performance and requirements. However, when the AI was reprogrammed with engineering-specific objectives, it began churning out innovative, high-performing designs.
The study highlights the need for a fundamental shift in how AI models are trained for engineering tasks. Focusing solely on statistical similarity can lead to overlooking essential design requirements, which can be detrimental in real-world applications. For example, two bicycle frames might look almost identical, but a minor difference might render one structurally inferior.
The team also experimented with a generative adversarial network (GAN) trained on thousands of bicycle frames. While the designs produced were realistic, they weren’t necessarily superior in terms of performance. In contrast, two other models designed specifically for engineering tasks resulted in better performing designs, although some lacked physical feasibility.
Regenwetter emphasized the need for AI models that prioritize design requirements over mere dataset similarity. He believes that AI can truly excel in design, but only if trained explicitly for the task.
The team’s findings, which were published in the journal Computer Aided Design, are a collaboration between computer scientists at the MIT-IBM Watson AI Lab and mechanical engineers from MIT’s DeCoDe Lab. The research underscores the potential of AI in revolutionizing engineering design, but also highlights the importance of training objectives that go beyond mere replication.
Assistant Professor Faez Ahmed envisions a future where AI models, with the right training, could greatly benefit various engineering domains. The team’s work is a stepping stone in redefining how generative AI is used in design, emphasizing innovation over replication.
Image: credit Faez Ahmed, Lyle Regenwetter, et al.