Using machine learning to infer rules for designing complex mechanical metamaterials
Mechanical metamaterials are artificial structures with mechanical properties that are driven by their structure and These structures have proved to be very promising for the development of new technologies
The University of Amsterdam have recently demonstrated the potential of convolutional neural networks, a class of machine learning algorithms, for designing complex mechanical metamaterials
Physical Review Letters, introduces two different CNN-based methods that can be used to derive and capture subtle combinatorial rules underpinning the design of mechanical metamaterials
As part of their previous research, van Mastrigt and his colleagues tried to distill the rules underpinning the successful design of complex meta materials
This was not an easy task, as the "building blocks" that make up these structures can be rearranged and altered in many different ways.
Simulation of all the ways in which building blocks can be rearranged using conventional physics simulation tools is possible when metamaterials have small cell sizes
conventional tools failed at allowing us to explore larger unit cell designs in an efficient way, the task becomes extremely challenging or impossible as these unit cell sizes become larger
Van Mastrigt and his colleagues initially had to overcome a number of challenges to successfully train CNNs to tackle the design of complex metamaterials
In the case of metamaterial M2, we attempted to create a training set that is class-balanced. For metamaterial M1, we added a reweight term to the loss function
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 much larger designs
The team found that they each performed better on different tasks, depending on the initial dataset used and known (or unknown) design symmetries.
the networks they trained were so far applied to a few metamaterial structures, they could be used to create much more complex designs, which would be incredibly difficult to tackle using conventional physics simulation tools
The work by van Mastrigt and his colleagues shows the huge value of CNNs for tackling combinatorial problems, where there are many variables at play, and for finding an optimal solution that complies with all constraints in the set
machine learning doesn't always allow researchers to view the processes behind a prediction or outcome, it can still be very valuable for exploring the design space for metamaterials
We are considering machine learning methods that will rare designs that have the properties that we want, if no examples of such designs are shown to the machine learning method before.