Engineering Design: Generative AI’s Journey Toward Innovation

Written By: Aniqa Batool
Last Updated On: October 26, 2023

Deep generative models when used to solve engineering design challenges. The team demonstrates that these models end up generating novel frames that replicate prior designs but fall short of engineering performance and criteria.

ChatGPT and deep generative models have shown to be amazing impersonators. By independently learning from past works, these AI supermodels are capable of cranking out poems, completing symphonies, and producing new movies and photographs. These extremely powerful tools succeed at creating new content.
MIT researchers point out in a new study, a similarity isn’t enough to effectively create engineering tasks.

A study author Explains that MIT mechanical engineering graduate student Lyle Regenwetter’s Deep generative models (DGMs) are very exciting, but they are also essentially flawed. “The goal of these models is to simulate a dataset. However, as engineers and designers, they frequently do not want to produce a design that is currently available.

He and his colleagues suggest that if mechanical engineers want generative AI to help them produce new ideas and solutions, they must first refocus the models beyond statistical similarity.

When the researchers created the same bicycle frame challenge to DGMs that were particularly constructed with engineering-focused aims in mind, instead of just numerical resemblance these models produced more creative and higher-performing frames.

In this paper, they describe the primary problems in applying DGMs to engineering design jobs and demonstrate that the underlying goal of typical DGMs is not individual design requirements into account. To exemplify this, the team used a basic scenario of bicycle frame design to show how difficulties might arise. As a model learns from thousands of existing bike frames of various sizes and shapes, it may be believed that two frames of similar dimensions have similar performance.