Algorithms that design structures have produced parts for ultralight aircraft, antennas and Formula 1 cars, but engineers who design buildings and bridges have been slow to trust them. MIT researchers think part of the reason is that the engineer only sees the result at the end. Their new system lets a designer step in while the algorithm is still working, and it costs very little extra time.
The technique is topology optimisation, which decides where material should go inside a given space to meet a set of loads. Normally it runs as a black box: you set the problem, wait, and get a finished shape that may look strange or be hard to build. If you want changes, you often start again. The MIT system, described in an open-access paper in Communications Engineering, pauses at points during the run and highlights regions of the design for the engineer to adjust.
The machine learning part decides which regions to show. Earlier attempts to add human judgement slowed the process badly, so the team trained a model to predict where a person would most likely want to intervene. To do that they built a synthetic dataset of probable human feedback on thousands of designs, covering appearance, performance and how easy a part would be to manufacture. In demonstrations, the human-guided designs were easier to make and had a 39 percent higher linear buckling load, while adding only 15 seconds to total design time compared with fully automated runs.
This is a different idea of an AI copilot from the coding assistants most readers know. Instead of the model producing an answer for a person to accept or reject, the person corrects the model mid-course, and the model learns where those corrections are likely. Senior author Josephine Carstensen argues there is room for both fully automated and collaborative approaches, depending on how much computing power a team has and how much human sign-off the work needs.
The limits are clear from the announcement. For now a designer can only make regions thicker or thinner; other kinds of change are planned for later versions. The feedback used for training is synthetic, not collected from practising engineers, so whether the model's guesses match what real designers want is still open. The results come from demonstrations in the paper rather than from structures built and tested in the field.