Dutch artificial intelligence researchers propose a method based on machine learning to design new complex metamaterials with useful properties

Combinatorial problems often arise in puzzles, origami, and metamaterial design. Such problems have rare sets of solutions that generate complex and distinct boundaries in the configuration space. Using standard statistical and numerical techniques, capturing these limits is often quite difficult. Is it possible to flatten a 3D origami piece without damaging it? This question is one of those combinatorial questions. As each crease must be consistent with the flattening, such results are difficult to predict just by looking at the design. To answer these questions, the UvA Institute of Physics and the AMOLF research center have shown that researchers can more efficiently and accurately answer such queries using machine learning techniques.

Despite using highly undersampled training sets, convolutional neural networks (CNNs) can learn to distinguish these boundaries for metamaterials in fine detail. This raises the possibility of complex hardware design by indicating that the network infers the underlying combinatorial rules from the sparse training set. The research team believes this will facilitate the development of sophisticated and functional metamaterials with artificial intelligence. The team’s recent study examined the accuracy of predicting the characteristics of these combinatorial mechanical metamaterials using artificial intelligence. Their work has also been published in the publication Physical Review Letters.

The attributes of man-made materials, which are engineered materials, are governed by their geometric structure rather than their chemical composition. Origami is one of these metamaterials. The ability of an origami piece to flatten is governed by how it is folded, i.e. its structure, not by the type of paper it is made from. More generally, intelligent design allows us to precisely regulate the bending, buckling or bulging of a metamaterial. This can be used for many different things, from satellite solar panels that deploy to shock absorbers.

A combinatorial metamaterial usually consists of two or more different orientations of building parts. These building blocks deform differently in response to an external mechanical force. The material will generally not yield under pressure if these building pieces are randomly mixed, as not all of them will be able to deform as desired. A neighboring building block should be able to taper inward where one wants to protrude outward. All deformed building components should fit together like a puzzle for the metamaterial to deform quickly. A “floppy” metamaterial can become rigid by modifying a single block, just as modifying a single fold can make an origami piece unflattened.

Although metamaterials have a wide range of potential uses, creating a new one is difficult due to their unpredictable behavior. It is usually trial and error to determine the general properties of metamaterials of different structures from a specific set of building blocks. Recent technological developments make it unnecessary for researchers to do all this work by hand. However, typical statistical and numerical methods are slow and error-prone because the properties of combinatorial metamaterials are very sensitive to changes in individual building blocks. This is where machine learning comes in. CNNs can accurately predict metamaterial properties of any building block configuration down to the smallest detail, even with only a minimal set of examples from from which to learn.

CNN’s results were astonishing and far beyond expectations. The accuracy of the predictions showed that the neural networks had mastered the fundamental mathematical principles governing the behavior of metamaterials, which still need to be better understood by the researchers themselves. These results indicate that complex metamaterials with relevant properties can be created using AI. More generally, applications of neural networks can help researchers solve combinatorial problems in various contexts and raise various intriguing concerns. The results can also help understand neural networks by showing how the complexity of a neural network correlates with the complexity of the problems it can handle.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Machine Learning of Implicit Combinatorial Rules in Mechanical Metamaterials'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and reference article.
Please Don't Forget To Join Our ML Subreddit


Khushboo Gupta is an intern consultant at MarktechPost. She is currently pursuing her B.Tech from Indian Institute of Technology (IIT), Goa. She is passionate about the fields of machine learning, natural language processing and web development. She likes to learn more about the technical field by participating in several challenges.


James G. Williams