The research on this matter was conducted by scientists from the universities of Amsterdam and Utrecht, as well as representatives of the AMOLF Institute. In the course of their activities, they used the so-called traditional neural networks related to machine learning.
Authors write about their accomplishments in the pages Physical review letters. As it turns out, they used two different approaches that build on traditional neural networks that can derive and capture the combinatorial rules underlying metamaterials design.
Our latest study can be seen as a continuation of the combinatorial design approach introduced in previous work, which can be applied to more complex building blocks. At the time I started working on this problem, Alexei Bossart and David Dykstra were working on a combinatorial metamaterial, which is multifunctional, meaning a material that can deform in many different ways depending on how it’s processed. launches.
Ryan van Maastricht explains
Metamaterials have properties that depend on their structure on a larger scale than molecules
One of the first challenges in achieving this goal was finding a way to effectively represent metamaterials designs. In the end, the researchers decided to use pixels. The idea is to encode the direction of each block in a clear visual way. The next step was related to the fact that there are many known metamaterials belonging to Class I, but much less than belonging to Class C, and it is this class that is of greatest interest.
The two methods proposed by the researchers yielded very promising results. Both approaches performed better than previous use. The researchers emphasize, however, that neural networks don’t just invade the design space. They can also find structure in this design space. In the near future, the advances made so far could help create more complex designs that lead to the formation of metamaterials with desired properties.
In our next research, we will pay attention to the reverse design. The current tool already helps us a lot in reducing the design space to find suitable (C-class) designs, but it doesn’t find us the best design for the task we have in mind. We are now considering using machine learning methods to help us find extremely rare constructs that have the properties we want, ideally even if no examples of such constructs have been previously shown to machine learning.
Van Maastricht concludes
“Infuriatingly humble musicaholic. Problem solver. Reader. Hardcore writer. Alcohol evangelist.”