The Pulse
MIT researchers test AI copilot for structural design feedback
The system recommends parts of an evolving structural design for an engineer to modify. In demonstration examples, the researchers report improved manufacturability or a 39% increase in linear buckling load.
A structural design system developed by MIT researchers uses machine learning to suggest parts of an evolving design where an engineer might want to make changes. In demonstration examples, the researchers report improved manufacturability or a 39% increase in linear buckling load, with 15 seconds added to total design time compared with conventional simplistic topology optimization.
The research appeared in the open-access journal Communications Engineering on November 10. The journal describes the article as an early version of peer-reviewed, accepted research that is subject to further editing before the final version of record.
Putting a designer in the loop
The tool works with topology optimization, a technique that iteratively fills the space available for a design to meet specified goals. Conventional approaches can take substantial computing time, and engineers may not see a completed design until the process is over, according to MIT News.
Some algorithm-generated designs can be difficult to build. MIT News gives the example of a design with eight or more beam connections at a single point; changing a finished design can mean starting the process again.
Lead author Dat Ha is a recent MIT civil and environmental engineering doctoral student and now a structural designer at LeMessurier. He told MIT News that complex designs need human input to make them more practical for industry.
Senior author Josephine Carstensen, an MIT associate professor of civil and environmental engineering, described the approach as a design copilot. She said engineers have had little influence over how topology-optimized designs develop, limiting their ability to control unwanted features.
Training on simulated preferences
The machine-learning model treats the task as image segmentation and uses a U-Net architecture. Its job is to predict regions that a user might prefer to modify and present them as recommendations while the design is being generated.
The model was trained on synthetic datasets representing preferences for either the longest structural member or the most complex structural connection. MIT News says the researchers generated feedback on thousands of designs, covering considerations such as appearance, performance and manufacturability.
The paper reports that the synthetic preference model generalized across varied and non-standard topology optimization problems. It also showed behavior beyond the single-region selections represented in its training data, the authors say.
Results and limits
The reported 15-second increase in total design time was measured against conventional simplistic topology optimization. The paper describes manufacturability and linear buckling load as alternative improvements in its demonstration examples, rather than saying every design achieved both.
For now, users can make structures thicker or thinner within identified regions. The researchers hope to add other modification options in future versions, according to MIT News.
The other authors are Ferdous Alam, a former MIT postdoctoral researcher now an assistant professor at Georgia Tech; Markus Buehler, an MIT professor of engineering; and Faez Ahmed, an MIT associate professor of mechanical engineering. The work received partial support from the MathWorks Engineering Fellowship Fund, TATA Steel and the MIT Generative AI Impact Consortium.