As I have said before, designers and engineers today should study origami and kirigami. These ancient arts embody a wealth of strategies for creating 3D structures from simple 2D materials. These techniques have inspired designs for flat pack furniture and spacecraft, for small robots, and for wearable electronics. (and this, this, this, and so on)
For a designer, ‘paper folding’ offers a gigantic design space of geometry and material physics. Learning how to fold a crane isn’t really helpful, we need to have CAD systems that understand the principles hidden in the old crafts, and how they work with materials other than paper.
And this is happening! In the past decade we have seen the emergence of “computational origami”, including the awkwardly named “Origamizer”.
This spring researchers at US Argonne National Lab report a tool for applying Kirigami techniques to create stretchable electronics, e.g., for wearable devices [3]. Kirigami includes techniques for making paper stretch as much as 50% with strategically placed cuts. For any given piece, there are millions of possibilities, and the results depend on the material. Brute force searching by trial and error isn’t feasible, especially for expensive and tricky materials like thin films of exotic materials.
In the early twenty first century, we solve problems with machine learning. It’s the hammer of the month.
In two papers the ANL researchers report a machine learning system for designing quantum materials [1, 2] (here and here). (I’ll note that this one journal alone, Computational Materials, has dozens of articles in recent years that discuss Kirigami-inspired material fabrication and design. This is definitely a hammer of the month!)
There are actually two parts to this design technique. First, there is a model that predicts the behavior of a sheet with different patterns of cuts. The second is a model of synthesizing the thin film by vapor deposition. The former identifies the best designs, the latter defines how to correctly and efficiently fabricate them.
These techniques were developed using reinforcement learning.
Simulating the behavior of these materials is slow, several hours at least for each candidate. So the system simulated a small random sample of the whole space, and used machine learning to develop a model that predicts the results.
Similarly, simulation of each synthesis schedule is far too complicated. The researchers trained a model using a sample of 10000 simulations. The model is able to successfully predict optimal processes for a given target.
Cool!
I’m not expert enough to know how general these results are. I suspect that you’d need to run the same learning for other materials and fabrication processes. Even so, the payoff is obvious: once you have the model in hand, you can design anything you want, as many times as you want.
As the researchers comment, these techniques can substantially shorten the time to develop new materials and new applications of materials, which is huge.
“Rapid development of technology based on advanced materials requires us to considerably shorten the existing ~20-year materials development timeline.
([1], p.1 )
It is easy to see that the next generation of CAD systems are going to have libraries of expert systems like these.
- Pankaj Rajak, Aravind Krishnamoorthy, Ankit Mishra, Rajiv Kalia, Aiichiro Nakano, and Priya Vashishta, Autonomous reinforcement learning agent for chemical vapor deposition synthesis of quantum materials. npj Computational Materials, 7 (1):108, 2021/07/14 2021. https://doi.org/10.1038/s41524-021-00535-3
- Pankaj Rajak, Beibei Wang, Ken-ichi Nomura, Ye Luo, Aiichiro Nakano, Rajiv Kalia, and Priya Vashishta, Autonomous reinforcement learning agent for stretchable kirigami design of 2D materials. npj Computational Materials, 7 (1):102, 2021/07/09 2021. https://doi.org/10.1038/s41524-021-00572-y
- John Spizzirri, Ancient art meets AI for better materials design, in Argonne National Laboratory – News, April 7, 2022. https://www.anl.gov/article/ancient-art-meets-ai-for-better-materials-design
Robert McGrath's Blog 