Designing complex mechanical metamaterials using machine learning algorithms

A research team from the University of Amsterdam and Utrecht University have proposed the use of ML (machine learning) to infer rules for designing the complex structures of mechanical metamaterials. The research was carried out by Ryan van Mastrigt, Marjolein Dijkstra, Martin van Hecke, and Corentin Coulais.

Mechanical metamaterials are artificially created materials designed to have specific and unique properties that don’t usually occur in nature. The range of these properties can be wide depending on the specific design and microstructure of the material.

Metamaterial array: copper wires on interlocking sheets of fiberglass circuit board. Source: Wikimedia Commons

Their article published on arXiv and Physical Review Letters proposes two distinct techniques based on Convolutional Neural Networks (CNNs). The two CNNs were specifically designed and trained to deduce the intricate combinatorial principles for the creation of mechanical metamaterials.

These metamaterials have unique mechanical properties, such as stiffness, strength, and damping. These properties are driven by their structure, rather than their composition. To achieve these amazing features they must often be structured at the nano or micro scale.

But designing these complex structures is difficult and can be hardly solved using standard statistical and numerical techniques.

The research team has recently shown that the CNNs have the ability to construct intricate mechanical metamaterials. The picture below shows two new combinatorial mechanical metamaterials designed by CNNs in such a way to protrude the letters “M” and “L” when they are pressed together between two plates.

Image source: Two complex mechanical metamaterials

Previous studies have demonstrated that it is possible to simulate every potential configuration and deformation of metamaterials when the size of their unit cell is tiny. But when size increases, the process becomes either impossible or very difficult.

The research team successfully trained two types of CNNs to develop more complex designs of metamaterials.

Firstly, the research team had to create a training dataset using the pixel representation of their metamaterial designs.

After training the two CNNs, they were able to create new complex designs, a task that would be very challenging for the conventional physics modeling tools.


The potential of CNNs for designing complex mechanical metamaterials may have future effects on how metamaterials are created. 

Additionally, these CNNs may be employed to develop even more complex designs in the future.

Learn more:

Other popular posts