Materials made of mechanical neural networks can learn to adapt their physical properties

A new type of material can learn and improve its ability to cope with unexpected forces through a unique lattice structure with connections of varying stiffness, such as described in a new article by my colleagues and me.

The new material is a type of architectural material, which derives its properties primarily from the geometry and specific characteristics of its design rather than what it is made of. Take hook-and-loop fabric closures like Velcro, for example. It does not matter if it is made of cotton, plastic or any other substance. As long as one side is fabric with stiff hooks and the other side has soft loops, the material will have the sticky properties of Velcro.

My colleagues and I based the architecture of our new material on that of an artificial neural network – layers of interconnected nodes that can learn to do chores by changing the importance or weight they give to each connection. We hypothesized that a mechanical network with physical nodes could be trained to take on certain mechanical properties by adjusting the stiffness of each connection.

Architected materials, like this 3D lattice, derive their properties not from what they are made of, but from their structure. Image credit: Ryan Lee, CC BY-ND

To find out if a mechanical network would be able to take on and retain new properties, such as taking on a new shape or changing directional strength, we started by building a computer model. We then selected a desired shape for the material along with the input forces and had a computer algorithm adjust the tensions of the connections so that the input forces produced the desired shape. We did this training on 200 different truss structures and found that a triangular truss was the best for getting all the shapes we tested.

Once the many connections are tuned to perform a set of tasks, the hardware will continue to react in the desired way. The formation is – in a sense – remembered in the structure of the material itself.

We then built a physical prototype array with adjustable electromechanical springs arranged in a triangular array. The prototype is made up of 6 inch connections and is about 2 feet long by 1 ½ feet wide. And it worked. When the network and the algorithm worked together, the material was able to learn and change shape in particular ways when subjected to different forces. We call this new material a mechanical neural network.

A photo of hydraulic springs arranged in a triangular lattice
The prototype is 2D, but a 3D version of this material could have many uses. Image credit: Jonathan Hopkins, CC BY-ND

why is it important

Besides some living tissue, very few materials can learn to handle unexpected loads better. Imagine an airplane wing that suddenly catches a gust of wind and is forced in an unexpected direction. The wing cannot change its design to be stronger in this direction.

The prototype network material we designed can adapt to changing or unknown conditions. In a wing, for example, these changes could be the accumulation of internal damage, changes in how the wing attaches to gear, or fluctuating external loads. Each time a wing made of a mechanical neural network experienced one of these scenarios, it could tighten and loosen its connections to maintain desired attributes like directional strength. Over time, as the algorithm makes successive adjustments, the wing adopts and maintains new properties, adding each behavior to the rest like a kind of muscle memory.

This type of material could have far-reaching applications for the longevity and efficiency of built structures. Not only could a wing made of a mechanical neural network material be stronger, but it could also be shaped to transform into shapes that maximize energy efficiency in response to changing conditions around it.

What is not yet known

So far, our team has only worked with 2D networks. But using computer modeling, we predict that 3D networks would have a much greater ability to learn and adapt. This increase is due to the fact that a 3D structure could have dozens of times as many connections, or springs, that do not intersect. However, the mechanisms we used in our first model are far too complex to be supported in a large 3D structure.

And after

The hardware that my colleagues and I have created is a proof of concept and shows the potential of mechanical neural networks. But bringing this idea to the real world will require figuring out how to make individual parts smaller and with precise bending and tension properties.

We hope that further research into the fabrication of micron-scale materialsas well as work on new materials with adjustable stiffnesswill lead to advancements that will make powerful intelligent mechanical neural networks with micron-scale elements and dense 3D connections a ubiquitous reality in the near future.

This article is republished from The conversation under Creative Commons license. Read it original article.

Image credit: Flexible Research Group at UCLA

Comments are closed.