Algorithms already help create parts for ultralight aircraft, broadband antennas and Formula 1 race cars. However, the most advanced algorithmic design methods require a lot of computation and time: the engineer sees only the finished result, and to change something it is often necessary to restart the process. Such designs can also be hard to verify for strength and to actually manufacture.
A group of researchers from the Massachusetts Institute of Technology (MIT) proposed making algorithmic design more collaborative. They developed a system that works like an AI partner: while the algorithm builds a structure, it shows the person individual regions and waits for feedback. An article about this was published in the journal Communications Engineering in open access.
The approach is based on topology optimization - a method in which an algorithm step by step redistributes material within a given space to best achieve the set goals. Usually such a process requires large computational resources, and the finished result may be inconvenient to manufacture. For example, eight or more beams may converge at one point - in reality such a joint is difficult to fabricate.
To give the engineer more control, the authors used machine learning. The model quickly identifies areas of the design that the person is likely to want to change. To train it, the researchers created a synthetic dataset - probable feedback on thousands of designs, where aesthetics, performance and manufacturability were evaluated. In other words, the algorithm learns to anticipate typical human edits.
In demonstration tests, the new approach improved manufacturability and linear buckling load (an indicator of how resistant a structure is to bending under compression) by 39% compared with a purely algorithmic method. At the same time, total design time increased only slightly - by just 15 seconds.
For now the system allows the person only to make selected regions thicker or thinner. In future versions the researchers hope to add other editing options. They also emphasize that the new approach does not replace fully automatic design but complements it: if a discipline has a lot of computational resources, a purely algorithmic method can also be used, and if the human role matters, the collaborative one.
According to the authors, human participation increases trust in the result: in some situations it is the engineer who must make the final decision before a design goes into production. The researchers believe their method will make topology optimization more accessible to a wider range of designers - not only in industry but also in academia.