Algorithmic design can produce structures optimized for goals such as strength and low weight, but engineers may have little say in how those designs take shape. A system developed by MIT researchers gives designers a chance to steer the process as it runs, aiming to make its results easier to build and trust.
Why the final design can be hard to use
The system works with topology optimization, a method that fills a defined design space through repeated calculations to meet specified goals. It has helped create parts for ultralight aircraft, high-bandwidth antennas, and F1 racing cars.
Those calculations can take substantial time and computing power. In conventional workflows, a designer may see the result only after the process ends. If the structure has an unwanted feature, making a change can mean starting the process over.
A design can also perform well on paper while posing practical manufacturing challenges. For example, a structure may have eight or more beams meeting at one point, a configuration that can be difficult to make. Such results can make engineers wary of relying on a process they cannot readily influence as it unfolds.
A copilot that flags places for feedback
The MIT system brings a human into the optimization process before the structure is complete. As the algorithm generates a design, the system identifies regions where a designer is likely to want a change and presents them for feedback.
To make that interaction quick, the researchers trained a machine-learning model using a synthetic dataset. It represented likely human feedback on thousands of designs, including judgments about aesthetics, performance criteria, and manufacturability.
In the current version, designers can make structures thicker or thinner in the regions identified by the system. That gives engineers a way to apply their experience while the design is evolving, rather than accepting or rejecting a finished result without input.
Performance gains with limited added time
In demonstrations, the human-in-the-loop method improved manufacturability and linear buckling load by 39 percent. The researchers reported that it added only 15 seconds to total design time compared with approaches relying entirely on algorithms.
The results point to a practical tradeoff: a small addition of time can make room for human judgment, with improvements to both a structure’s manufacturability and its ability to resist buckling. The researchers see that combination as useful where designers need to assess whether an algorithmic result makes sense in practice.
Topology optimization has gained traction in automotive and aerospace design, but it has been used less for buildings and bridges. The researchers say that limited designer control and ownership can hold back wider use. Inviting engineers to guide a design may help address those concerns by making the process more collaborative and its results more acceptable to the people expected to build them.
A role alongside fully algorithmic design
The researchers present the copilot as an option that can complement AI-only design methods. Some fields may have the computing resources to rely more heavily on algorithms; in situations where human input is needed, a collaborative workflow can provide a way to incorporate it.
The approach is still limited in how users can modify a structure. The team hopes to add more kinds of changes beyond thickening or thinning selected regions. If those options expand, topology optimization could become accessible to a broader range of design engineers, as well as researchers and students.
The work was supported, in part, by the MathWorks Engineering Fellowship Fund, TATA Steel, and the MIT Generative AI Impact Consortium.