Complex artificial structures called mechanical metamaterials have mechanical properties that are determined by their structure rather than their content. Although the development of new technology has shown these structures to be very promising, designing them can be difficult and time-consuming.
Convolutional neural networks (CNNs), a family of machine learning algorithms, have recently been shown to have the potential for building intricate mechanical metamaterials by researchers from the University of Amsterdam, AMOLF, and Utrecht University. Their article, which was released in Physical Review Letters, proposes two distinct CNN-based techniques that can identify and capture the fine-grained combinatorial principles underlying the development of mechanical metamaterials.
According to Ryan van Mastrigt, one of the study’s authors, their current study can be considered a continuation of the combinatorial design approach outlined in a prior paper, which can be applied to more intricate building blocks. Aleksi Bossart and David Dykstra were working on a combinatorial metamaterial that is able to host many functions, meaning a material that can deform in numerous distinct ways depending on how one actuates it.
Van Mastrigt and his associates attempted to isolate the principles governing the effective creation of complicated metamaterials as part of their earlier research. As the “building blocks” that make up these structures can be distorted and placed in an almost infinite number of different ways, they rapidly recognized that this was by no means an easy undertaking.
Previous research demonstrated that it is possible to simulate every potential configuration and deformation of metamaterials when their unit cell sizes are tiny (i.e., when there are few “building blocks” available). But when these unit cell-sizes increase, the work becomes either impossible or very difficult.
They chose to seriously examine machine learning because they were unable to reason about any underlying design rules and conventional methods were ineffective at enabling them to efficiently explore larger unit cell designs, according to van Mastrigt. The primary goal of their research was to find a machine learning technique that would speed up their exploration of the design space. With their findings, he believes they were successful and even surpassed their own expectations.
Van Mastrigt and his colleagues first had to overcome a number of obstacles in order to successfully train CNNs to handle the design of complicated metamaterials. They initially had to figure out how to represent their metamaterial designs with accuracy.
According to van Mastrigt, they tried several approaches before settling on the pixel format. This representation encodes each construction block’s orientation in a plain visual manner, reducing the classification challenge to a visual pattern recognition problem, which is exactly what CNNs are good at, the author writes.
The researchers then had to come up with strategies that took into account the enormous class-imbalance in metamaterials. To put it another way, training CNNs to infer combinatorial rules for the various classes can entail different processes because there are now more known metamaterials belonging to class I than class C, which is the class that the researchers are interested in.
To solve this issue, Van Mastrigt and his colleagues devised two distinct CNN-based techniques. For different sorts of metamaterial, categorization problems can be resolved using these two techniques.
Van Mastrigt stated, In the instance of metamaterial M2, they aimed to develop a training set that is class-balanced. They accomplished this via naive undersampling (i.e., discarding a large number of class I samples) and combine this with symmetries, such as translational and rotational symmetry, which they know some designs have, to generate more class C designs.
As a result, some domain knowledge is required for this method. They introduced a reweight term to the loss function for metamaterial M1 so that the uncommon class C designs weigh more heavily during training, with the important notion being that this reweighting of class C cancels out with the far higher number of class I designs in the training set. This method necessitates no domain expertise.
Both of these CNN-based techniques for determining the combinatorial principles underlying the creation of mechanical metamaterials showed extremely encouraging results in preliminary tests. The researchers discovered that based on the baseline dataset utilized and known (or unknown) design symmetries, they each performed better on different tasks.
Van Mastrigt stated, they demonstrated exactly how exceptionally good these networks are at handling complex combinatorial tasks. This came as a huge surprise to them because all other traditional (statistical) approaches that physicists often utilize to solve similar issues are ineffective. They demonstrated that neural networks actually do more than simply extrapolate the design space depending on the examples you provide them with, since they seem to be strangely biased to create a structure (which derives from rules) in this design space that generalizes very well.
These researchers’ most recent discoveries may have profound effects on how metamaterials are created. The networks they trained have been used so far to build a few metamaterial structures, but they might also be utilized in the future to build much more intricate designs that would be very challenging to implement using conventional physics simulation tools.
Van Mastrigt and colleagues’ research also reveals the immense importance of CNNs for dealing with combinatorial issues, which are optimization tasks that necessitate the creation of an “optimal object” or arriving at a “optimal solution” that satisfies all constraints in a set, in situations where there are many variables at play. This article could encourage the usage of CNNs in different research and development environments as combinatorial challenges are prevalent in many scientific domains.
The researchers demonstrated that machine learning can still be very helpful for exploring the design space for metamaterials, as well as perhaps other materials, objects, or chemical substances, even though it is typically a “black box” approach and does not always allow researchers to view the processes behind a given prediction or outcome. This might thus make it easier to analyze and comprehend the intricate laws that underlie successful designs.
In their upcoming research, they will focus on inverse design, Added van Mastrigt. Although the existing tool greatly aids in reducing the design space in order to locate suitable (class C) designs, it does not assist them find the optimum solution for the task at hand. They are now researching machine learning techniques that will enable them to locate exceedingly unusual designs with the desired attributes, ideally even when the machine learning technique has never seen examples of such designs before.
Because of our recent research, they are optimistic that neural networks will help them solve this challenging problem.
