ChatGPT and other deep generative models are impressive mimics. These AI models can create poems, finish symphonies, and generate videos and images by learning from millions of examples. However, MIT engineers argue that these models are flawed because their objective is to mimic existing designs, which limits their ability to truly innovate in engineering tasks.
The researchers believe that if mechanical engineers want AI to help them generate novel ideas and designs, they need to refocus these models beyond statistical similarity. Statistical similarity alone is not enough for innovation in design. The models need to be different and unique.
In their study, the researchers show the limitations of deep generative models when applied to engineering design problems. They use a case study of bicycle frame design to demonstrate that these models generate new frames that mimic existing designs but fail in terms of engineering performance and requirements.
However, when the researchers designed AI models specifically for engineering-focused objectives, these models produced more innovative and higher-performing frames. This suggests that AI models with a focus on design requirements can be effective in assisting engineers in creating innovative products.
The researchers conclude that AI models can be valuable co-pilots for engineers if they are trained with task-appropriate metrics and priorities. By understanding the requirements and constraints of a design task, AI can help engineers be better and faster at creating innovative products.
The study, conducted by computer scientists at MIT-IBM Watson AI Lab and mechanical engineers in MIT’s DeCoDe Lab, highlights the need to go beyond statistical similarity in deep generative models. The research was published online and will appear in the December print edition of the journal Computer Aided Design.
Deep generative models, like ChatGPT, are powerful learners that can process large amounts of data and generate realistic samples. However, they have been primarily used to mimic existing designs in various engineering domains. Designers have used these models to create new aircraft frames, metamaterial designs, and optimal geometries for bridges and cars. But, without focusing on design requirements, these models have not improved on existing designs.
The researchers demonstrate the limitations of standard deep generative models in engineering tasks using the example of bicycle frame design. They show that even in the initial learning phase, these models fail to consider specific design requirements. This results in generated designs that may visually resemble existing frames but perform poorly due to small disconnects in their structural integrity.
To illustrate this point, the researchers trained a conventional generative adversarial network (GAN) on a dataset of bicycle frames and asked it to generate new designs. The generated designs were similar to existing frames but did not show any significant improvement in performance. In fact, some designs were even inferior.
In contrast, when the researchers used AI models specifically designed for engineering tasks, they generated realistic designs that were not only visually appealing but also lighter and stronger than existing designs. These models took into account design constraints and prioritized physically feasible frames.
The researchers believe that going beyond statistical similarity in AI models can lead to better designs and innovation in various engineering fields. By training models with a focus on performance, design constraints, and novelty, AI can have a significant impact on molecular design, civil infrastructure, and other engineering domains.
The study sheds light on the limitations of relying solely on statistical similarity in generative AI applications. It emphasizes the importance of considering specific design requirements when using AI models to assist engineers in creating innovative products.