TL;DR: MIT researchers have published an open-access study showing an AI copilot can flag which regions of an algorithm-generated structure a person should adjust, and in their demonstrations the human-in-the-loop workflow either made a design easier to manufacture or raised its linear buckling load by 39 percent, at a cost of 15 extra seconds.
The research appeared in Communications Engineering on November 10, 2026, under the title "AI-guided human-in-the-loop inverse design of high performance engineering structures," and MIT News described the system on November 13, 2026. The paper is open access. The copilot is built for topology optimization, a design method that searches for the best distribution of material inside a given space.
Dat Quoc Ha is the lead author and a recent MIT civil and environmental engineering doctoral student who now works as a structural designer at LeMessurier, according to MIT News. Josephine V. Carstensen, an MIT associate professor of civil and environmental engineering, is the senior author. Their co-authors are Md Ferdous Alam, a former MIT postdoc who is now an assistant professor at Georgia Tech; Markus J. Buehler, who holds MIT's Jerry McAfee (1940) Professorship of Engineering; and Faez Ahmed, an MIT associate professor of mechanical engineering. The paper lists Ha as its corresponding author.
The setup: engineers define the problem, then wait
A topology optimization run starts with the engineer defining the physical space that material can occupy, along with the anticipated loads and boundary conditions, the authors write in the paper's introduction. The design problem is written as an objective function plus constraints, and the algorithm searches for the material layout that best satisfies them, choosing which areas become solid and which stay empty. The authors' introduction says the field serves everything from aircraft components to offshore wind turbines and new lightweight architectures, each of which has to trade material efficiency against structural integrity.
Judging a candidate layout usually means dividing the space into elements and running a finite element analysis to measure a mechanical property, the authors write. Their description of the constraints covers equilibrium with the applied loads and a cap on how much material is available.
Quality is judged at the end. Once the algorithm returns a layout, the engineer inspects it and decides whether to redesign, which means resetting the problem and starting again. The authors describe this sequence as putting the user in what they call a retroactive judgment stage, and they call the technique a black box.
Those reruns add computation. MIT News notes that even the strongest algorithmic methods demand heavy computation and long run times, and that the resulting geometry can be awkward to build. One example the article gives is a design with eight or more beam connections meeting at a single point, a junction that would be difficult to produce. MIT News ranks topology optimization among the most advanced and promising algorithmic design methods, noting that it repeatedly redistributes material until a stated goal is met.
Interactive versions of topology optimization have tried to fill that gap, but they introduce their own delay. According to the paper, those methods depend on a person repeatedly choosing which region to modify, one trial after another. Previous attempts to keep human judgment in the loop also slowed the work dramatically, MIT News reports. Cutting the number of those trials is the point of the learned predictor.
How the copilot joins the loop
While the optimization is running, the model predicts the areas a designer would probably want to change and offers them as suggestions. The designer can make the structure inside those flagged regions thicker or thinner, and the algorithm resumes with that input. Instead of receiving a finished shape, the engineer shapes the search while it is still moving.
The model does not replace the optimizer. The algorithm still searches the design space for a material layout that meets the objective and constraints, while the predictions narrow where a designer is invited to step in. Carstensen compared the arrangement to working alongside a design copilot, and the paper places it within a wider move toward human-in-the-loop design, where a person's intuition enters the generation process.
This changes the usual routine, in which the finished geometry arrives only at the end of a run and users often start again to make adjustments, as MIT News describes it. Under the new loop, the human stays inside the same run rather than in a review that follows it.
The prediction model is set up as an image segmentation task and uses a U-Net architecture, according to the paper. It was trained on synthetic datasets in which the preferred region is defined in one of two ways: the longest member in the structural layout, or the most complex connection. The paper's abstract frames the task as predicting a synthetic user's preferred regions.
MIT News reports that the researchers created a synthetic dataset of probable human feedback covering thousands of designs, with critiques addressing aesthetics, performance criteria and manufacturability. Ha said the team reasoned that if a record existed of where people had altered these programs, there might be a pattern in those choices, similar to other machine-learning approaches.
The training preferences were generated rather than collected, with two rules standing in for a designer's judgment. The paper reports that the model predicted plausible regions for modification, handled a variety of topology optimization problems including non-standard ones, and showed behavior that went beyond the one-region-at-a-time examples it learned from.
What the demonstrations measured
In the paper's demonstration examples, adding the copilot either made a design easier to manufacture or raised linear buckling load 39 percent above a conventional, simplified topology optimization run, and the whole design run took 15 seconds longer. The comparison baseline is the standard automated workflow, which runs without a designer in the loop. MIT News summarized the outcome as structures that are easier to manufacture with higher buckling loads, at the cost of 15 seconds added to the run.
The paper's stated goal is to cut the number of iterative trials a designer needs. The figures come from examples the research team chose and ran, using a preference model trained on synthetic feedback rather than records of decisions by practicing engineers. The paper describes the framework as a scalable basis for real-time collaboration that stays aligned with user preferences in high-performance engineering.
For now, users can only make structures within identified regions thicker or thinner, and the researchers say they hope to add other modification options in later versions.
A full topology optimization run requires substantial computation, according to MIT News, and the paper measures the copilot's addition in seconds.
Why control and trust matter for buildings and bridges
Automotive and aerospace teams have picked up topology optimization; building and bridge designers have not, in large part because engineers have to be confident the result is sound, according to MIT News. Carstensen said a central limitation has been the limited say a design engineer has in how a design develops, which can leave unwanted features in place, and she said the shortage of control and ownership also holds the method back.
"Right now, the human doesn't have any agency or control over the design process as it's playing out," Ha said. He described the team's aim as combining algorithmic search with human judgment to make complex design problems more practical for industry.
Carstensen describes the system as blending the strengths of algorithmic and human design. "The idea is to tap into the brains of the designers and get them to infuse their domain knowledge as the design evolves," she said.
She also positions the copilot as a complement to fully automated design tools. "We are arguing there's room for both," she said, adding that fields with large computational resources can use the automated approach, while work that needs human input can use the collaborative one.
Trust is the reason for keeping a person in the loop, in her account. Carstensen said that in some settings people must sign off on the final design before it advances, and that involving them can bring higher-performing designs into the real world, because the aim is not simply good artifacts but artifacts people are willing to accept and build.
She said the hope is that clearing these hurdles brings topology optimization to far more users, in industry as well as among researchers and students in academia.
What the paper records
The published paper lists funding from the MathWorks Fellowship and the MIT Generative AI Consortium, and the authors report no competing interests. A preprint of the work was posted to arXiv on January 15, 2026 and is listed there at 21 pages with 10 figures.
The paper's introduction traces the black-box concern to earlier work. The authors write that previous studies have recommended giving designers a way to contribute their creativity, and that interviews with practicing designers have reinforced the point. They also write that the method frames design as a math problem and can yield efficient, lightweight structures without the biases of human intuition.
MIT News points to work on ultralight aircraft parts, high-bandwidth antennas and Formula 1 cars as examples of what algorithmic design has produced. The researchers argue that the copilot approach could put algorithmic design within reach of more engineers and make its output easier to manufacture, according to the same account.
What comes next
The authors say they hope future versions will allow modifications other than thickening and thinning, and they describe the current framework as a foundation for real-time collaboration in high-performance engineering. The paper concludes that AI copilots can remove iterative bottlenecks from interactive design.
The work was reported by MIT News on November 13, 2026.