MIT researchers improve structural designs with AI-guided human feedback

Engineers can help shape an algorithm-generated structure before it is finished, rather than wait for a final design that may be difficult to build. MIT researchers have demonstrated an AI-assisted method that recommends places to intervene during design. Their examples improved manufacturability or increased linear buckling load, while adding about 15 seconds to the process.

The work, published November 10 in Communications Engineering, addresses a bottleneck in topology optimization: an algorithmic technique that repeatedly adjusts where material goes within a design to meet specified goals. Human judgment can make the resulting structures more practical, but repeatedly choosing regions to modify takes time. The researchers use machine learning to help with that selection.

Finding the places an engineer would change

An optimized structure can be awkward to manufacture even when it meets the algorithm’s objectives. MIT gives the example of a design with eight or more beams meeting at one point—a connection that would be difficult to make. Waiting until optimization finishes to address such features can mean starting the design process again.

The new system instead presents suggested regions while the structure is evolving. Its recommendation model uses U-Net, a machine-learning architecture for identifying regions within images. Here, the task is to mark plausible areas for an engineer to modify, reducing the repeated region-selection work that has slowed earlier human-in-the-loop approaches.

The training preferences were synthetic: examples constructed to represent likely choices, rather than a record of engineers editing actual projects. The paper specifies two preferences used to select regions:

• Identify the longest structural member in a design.
• Identify the most complex structural connection.

The researchers report that the preference model generalized across varied and nonstandard optimization problems. It also displayed behavior beyond the single-region selections used in training.

The idea is to tap into the brains of the designers and get them to infuse their domain knowledge as the design evolves

Josephine Carstensen, MIT associate professor of civil and environmental engineering

A measured gain, not a universal percentage

The demonstration results distinguish two outcomes: better manufacturability, or a 39% improvement in linear buckling load, a structural-performance measure. The paper does not describe a universal 39% improvement in both measures. Its wording is narrower than MIT News’ summary, which joins manufacturability and buckling load with “and.”

The extra 15 seconds is also a comparison from the demonstrations, measured against what the authors call conventional, simplistic topology optimization.

Human control still has a narrow scope

For now, users can only make structures thicker or thinner within the identified regions. The model recommends where to intervene; the engineer supplies the modification. The team hopes to add other editing options in future versions. As described in MIT News on November 13, this is a research approach with a limited set of controls, rather than unrestricted interactive structural design.

Carstensen frames the collaborative method as a complement to AI-only design, not a replacement. She says different settings can support different approaches, depending on computational resources and the need for human input. Her argument also concerns acceptance: involving engineers during generation can give them more influence and ownership over structures they may ultimately have to approve and build.