An AI Co-Pilot Suggests Which Part of a Machine-Made Design to Change Next

A team at the Massachusetts Institute of Technology has added a step to structural design software that proposes which region of a computer-generated shape a person should rework. The work was published Nov. 10, 2026, in Communications Engineering, and the authors call the addition an AI co-pilot: it marks the candidate regions, and the person still makes the change.

The paper reports that topology optimization, software that lays out a structure from scratch to carry a load, is held back by long computing times and by a black-box quality that leaves the user nothing to interact with. Interactive versions that put a person back in the loop rely on a slow cycle of picking a region to modify, one trial at a time. Cutting the number of those trials is what the co-pilot is for.

The authors report that the prediction model is set up as an image-segmentation task, labeling which areas of a picture to act on, and built on a U-Net architecture. It was trained on synthetic datasets in which the preferred region is either the structure's longest member or its most complex connection, so the preferences it learned were generated rather than collected from engineers. The authors report that the model proposes plausible regions, presents them to the user as recommendations, generalizes to non-standard design problems, and behaves in ways its single-region training data did not cover.

On the team's own demonstration examples, the authors report that adding the co-pilot either made a design easier to manufacture or raised the linear buckling load, the force a part carries before it buckles, by 39%. Total design time rose by 15 seconds compared with conventional topology optimization. Both figures are the group's own, from examples it chose.

The journal published the accepted version early, with a transparent peer review file, and says a final version of record will replace it. Dat Quoc Ha and Josephine V. Carstensen of MIT led the work with colleagues in the university's mechanical engineering department. The MathWorks Fellowship and the MIT Generative AI Consortium funded it.