You can freely download images from the MIT News office website under a Creative Commons Attribution Non-Commercial No Derivatives license. These images are available to non-commercial entities, the press, and the general public. Please note that you are not allowed to modify the images provided, except for cropping them to the desired size. When reproducing the images, you must give credit by using a credit line. If none is provided, credit should be given to "MIT. " MIT engineers have trained several AI models on thousands of bicycle frames from a dataset of bicycle designs. These frames are color-coded based on bike style. Deep generative models like ChatGPT can mimic various creative works, including poems, symphonies, videos, and images, by learning from numerous examples. However, the researchers at MIT emphasize that simply mimicking existing designs is not enough for true innovation in engineering tasks. According to Lyle Regenwetter, a mechanical engineering graduate student at MIT and the author of the study, deep generative models are promising yet inherently flawed. The models aim to mimic a given dataset, but engineering and design require creating something that doesn't already exist. Therefore, to generate novel ideas and designs, mechanical engineers need to refocus these models beyond just statistical similarity. Faez Ahmed, assistant professor of mechanical engineering at MIT and co-author of the study, explains that the performance of many deep generative models is directly tied to how closely a generated sample resembles what the model has learned. However, in design, being different is important for true innovation. The researchers conducted a case study on bicycle frame design, demonstrating that when models were designed with engineering-focused objectives instead of solely relying on statistical similarity, they produced more innovative and higher-performing frames. The study by Ahmed and Regenwetter highlights the limitations of similarity-focused AI models when applied to engineering problems. However, they also suggest that with careful planning and task-appropriate metrics, AI models can be effective design "co-pilots. " The main objective is for AI to help engineers create innovative products more efficiently. The research, a collaboration between computer scientists at the MIT-IBM Watson AI Lab and mechanical engineers at MIT's DeCoDe Lab, is published online and will be featured in the December print edition of the journal Computer Aided Design.

The study's co-authors include Akash Srivastava and Dan Gutreund from the MIT-IBM Watson AI Lab. Deep generative models, often referred to as DGMs, are powerful learners that can process vast amounts of data. DGM is a broad term that encompasses machine-learning models trained to understand the distribution of data and generate new content that is statistically similar. ChatGPT, for example, is a popular large language model that can generate realistic imagery and speech based on conversational queries. Other models like DALL-E and Stable Diffusion are also widely used for image generation. While DGMs have been increasingly applied in various engineering domains, they have mostly replicated existing designs without improving performance. Designers often overlook the importance of adjusting the model's training objective to focus on specific design requirements. As a result, designs generated by these models are very similar to the existing dataset. In their study, the researchers identify the main challenges when applying DGMs to engineering tasks and demonstrate that the standard objective of these models does not consider the specific design requirements. They use bicycle frame design as a simple case to illustrate the potential problems that can arise. For instance, a DGM might consider two frames with similar dimensions to have similar performance, when in reality, a small difference could significantly affect the frame's strength. To test the performance of DGMs, they trained a conventional generative adversarial network (GAN) on a dataset of thousands of bicycle frames. The model successfully generated realistic designs resembling existing frames. However, these designs did not show significant improvements in performance and, in some cases, were even inferior. The researchers then tested two other specifically engineered DGMs for engineering tasks. The first model, developed by Ahmed, generated realistic designs that were both lighter and stronger than existing ones but also had physically "invalid" components. The last model, built by Regenwetter, generated the highest-performing designs that were also physically feasible. Ahmed believes that if DGMs are trained with priorities such as performance, design constraints, and novelty, various engineering fields could benefit greatly. By highlighting the limitations of relying solely on statistical similarity, the researchers hope to inspire new approaches and strategies for generative AI applications in fields beyond multimedia.