Ingrid Fadelli is a writer for the website Phys.org.
The mechanical metamaterials have mechanical properties that are driven by their structure. These structures are very promising for the development of new technologies, but they can be difficult to design and take a lot of time.
Researchers at University of Amsterdam, AMOLF, and Utrecht University have recently demonstrated the potential of convolutional neural networks. There are two different CNN-based methods that can be used to derive and capture the subtle combinatorial rules underpinning the design of mechanical metamaterials.
Ryan van Mastrigt, one of the researchers who carried out the study, said that the study can be considered a continuation of the previous paper. Around the time I started working on this study, David Dykstra and Aleksi Bossart were working on a metamaterial that can host multiple functions, meaning a material that can change in many different ways depending on how one actuates it.
The rules underpinning the successful design of complex metamaterials were distilled by van Mastrigt and his colleagues. This was not an easy task as the building blocks that make up these structures can be rearranged in many different ways.
Simulation of all the ways in which building blocks can be rearranged using conventional physics simulation tools can be done when metamaterials have small cell-sizes. The task becomes more difficult as the cell-sizes get bigger.
We decided to consider machine learning as a serious option since we were unable to reason about any underlying design rules and conventional tools failed at allowing us to explore larger unit cell designs in an efficient way. The main goal of our study was to find a machine learning tool that would allow us to explore the design space quicker than before. I believe that we succeeded and exceeded our expectations.
Van Mastrigt and his colleagues had to overcome a number of challenges in order to train CNNs to tackle the design of complex metamaterials. They had to find a way to represent their design.
"We tried a number of approaches and finally settled on what we refer to as thepixel representation," van Mastrigt said. "This representation shows the orientation of each building block in a clear way, so that the classification problem is cast to a visual pattern detection problem, which is exactly what CNN's are good at."
The researchers had to come up with ways to consider the huge metamaterials class-imbalance. As there are many known metamaterials belonging to class I, but not many belonging to class C, training CNNs to infer combinatorial rules for these different classes might be different.
Two CNN-based techniques were devised by van Mastrigt and his colleagues. Metamaterial classes and classification problems can be solved using these two techniques.
Van Mastrigt said that they tried to create a training set that was class balanced. We used nave undersampling and symmetries to create additional class C designs.
The approach needs some domain knowledge. For metamaterial M1, we added a reweight term to the loss function so that the rare class C designs weigh more heavily during training, where the key idea is that this reweighting of class C cancels out with the large number of class I designs. No domain knowledge is required.
The results of the initial tests were very promising. Depending on the initial dataset used and known design symmetries, the team performed better on some tasks.
Van Mastrigt said that they showed how good these networks are. The use of fail for these types of problems is common among physicists. Neural networks do more than just interpolate the design space based on the examples you give them, as they seem to be biased to find a structure that generalizes extremely well.
The findings gathered by this team of researchers could have far reaching implications. While the networks they trained were so far applied to a few metamaterial structures, they could eventually also be used to create far more complex designs, which would be very difficult to do using conventional physics simulation tools.
The work by van Mastrigt and his colleagues shows how CNNs can be used to tackle combinatorial problems, where there are many variables at play, and to find an optimal solution that complies with all constraints. CNNs can be used in other research and development settings as a result of the paper's recommendation.
Even if machine learning is a black box approach, it can still be useful for exploring the design space for metamaterials. This could help to understand the rules behind effective designs.
We will look at inverse design in our next studies. The current tool helps us reduce the design space, but it doesn't find us the best design for the job we have in mind. When no examples of such designs are shown to the machine learning method, we are considering machine learning methods that will help us find extremely rare designs that have the properties that we want.
We think that neural networks will allow us to successfully tackle the problem.
Ryan van Mastrigt and his team wrote Machine Learning of Implicit Combinatorial Rules in Mechanical Metamaterial. There is a book titled "PhysRevLett.129.198003."
Corentin Coulais and his team designed textured mechanical meta materials. Nature 18960 is published in the journal Nature.
Exotic mechanics are produced by topological defects in complex meta materials.
The Proceedings of the National Academy of Sciences has a paper on oligomodal meta materials. 10.1073/pnas.
There is a science network.
All Rights Reserved © 2024
