Two combinatorial mechanical metamaterials designed in such a way that the letters M and L bulge out in the front when being squeezed between two plates (top and bottom). Designing novel metamaterials such as this is made easy by AI. Credit: Daan Haver and Yao Du, University of AmsterdamA new study led by the University of Amsterdam and research institute AMOLF in The Netherlands has shown that artificial intelligence can predict the properties, and help in the design, of complex functional metamaterials. Machine-learning algorithms were found to predict the properties of combinatorial mechanical metamaterials, which can be controlled to accurately bend, buckle or bulge, a breakthrough that could find uses in shock absorbers, and even unfolding solar panels on space satellites.
A combinatorial metamaterial can be built up of two or more types or orientations of building blocks, which will deform in specific ways on a mechanical force being applied. When the blocks are combined randomly, they tend to jam and the material will not buckle under pressure because not all blocks are able to deform the way they want to. For a metamaterial to easily buckle, all deformed building blocks need to fit together like a jigsaw puzzle, and changing a single block can make a ‘floppy’ metamaterial rigid.
However, designing a metamaterial for a specific application is problematic, especially as the properties of combinatorial metamaterials are sensitive to changes to individual building blocks. This means the standard statistical and numerical approaches are slow and prone to error. Instead, as reported in Physical Review Letters [van Mastrigt et al. Phys. Rev. Lett. (2022) DOI: 10.1103/PhysRevLett.129.198003], this study demonstrated that neural networks can accurately predict the metamaterial properties of any configuration of building blocks, even down to the smallest detail.
The finding means it could be possible to use AI to design new complex metamaterials with useful properties, and maybe assist in solving combinatorial problems in other circumstances. As first author Ryan van Mastrigt told Materials Today, “The main significance of this work is that we show just how extraordinarily good these networks are at solving complex combinatorial problems. This is really surprising, since all other conventional (statistical) tools we as physicists are used to using fail for these types of problems”.
Combinatorial problems in areas such as origami and metamaterial design involve rare sets of solutions that define complex and sharply delineated boundaries in configuration space. The machine-learning tool, effectively a black box,can predict whether a particular set of mechanical building blocks fits together to yield a desired deformation, as it learns from test structures which blocks squeeze snugly together and then uses that information to make predictions about other potential structures.
Apart from mechanical metamaterials, the neural network tool could be used to find functional structures for any system built from blocks, such as proteins that fold and chemical compounds with biological functions. The team now hope to explore inverse design, as their approach helps to reduce the design space to find suitable designs.
“The main significance of this work is that we show just how extraordinarily good these networks are at solving complex combinatorial problems. This is really surprising, since all other conventional (statistical) tools we as physicists are used to using fail for these types of problems”.Ryan van Mastrigt