A mechanical neural network is a new type of material that can learn and change its physical properties

A mechanical neural network is a new type of material that can learn and change its physical properties. This lets it be used to make strong structures that can adapt to different situations.

The main concept

Because of a unique lattice structure with variable stiffness connections, a new type of material can learn and improve its ability to deal with unexpected forces, as described in a new paper by my colleagues and me. 

The new material is a type of architected material, which derives its properties primarily from the geometry and specific characteristics of its design rather than the material from which it is made. Consider hook-and-loop fabric closures such as Velcro. It makes no difference whether it is made of cotton, plastic, or another material. The material will have the sticky properties of Velcro as long as one side is a fabric with stiff hooks and the other side is a fluffy loop.

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 tasks by varying the importance, or weight, placed on each connection. We thought that a mechanical lattice with physical nodes could be taught to have certain mechanical properties by changing how rigid each connection was.

To determine whether a mechanical lattice could adopt and maintain new properties, such as changing shape or directional strength, we first created a computer model. We then chose a desired shape for the material as well as input forces, and a computer algorithm was used to tune the tensions of the connections so that the input forces produced the desired shape. We tested 200 different lattice structures and discovered that a triangular lattice was the best at achieving all of the shapes we tested.

Once the many connections have been tuned to accomplish a set of tasks, the material will continue to react as desired. In some ways, the training is remembered in the structure of the material itself.

We then created a physical prototype lattice with triangular electromechanical springs. The prototype is 2 feet long by 112 feet wide and made of 6-inch connections. And it was effective. When the lattice and algorithm collaborate, the material learns and changes shape in specific ways when subjected to different forces. This new material is known as a "mechanical neural network." 

Why is it significant?

Aside from living tissues, very few materials can learn to be more resilient to unexpected loads. Consider a plane wing that unexpectedly catches a gust of wind and is pushed in an unexpected direction. The wing's design cannot be changed to make it stronger in that direction.

We designed a prototype lattice material that can adapt to changing or unknown conditions. These changes in a wing could include the accumulation of internal damage, changes in how the wing is attached to a craft, or fluctuating external loads. When a mechanical neural network wing encountered one of these scenarios, it could strengthen and soften its connections to maintain desired attributes such as directional strength. As the algorithm is changed over and over, the wing takes on and keeps new properties. Each new behavior builds on the others, like muscle memory.

This type of material could have far-reaching applications in terms of building longevity and efficiency. A wing made of a mechanical neural network material could not only be stronger, but it could also be trained to morph into shapes that maximize fuel efficiency in response to changing environmental conditions.

What is still unknown?

Our team has only worked with 2D lattices so far. However, we predict that 3D lattices will have a much greater capacity for learning and adaptation based on computer modeling. This increase is due to the fact that a 3D structure may have tens of times more non-intersecting connections, or springs. The mechanisms we used in our first model, on the other hand, are far too complex to be supported in a large 3D structure.

What comes next?

My colleagues and I created a proof-of-concept material that demonstrates the potential of mechanical neural networks. To put this idea into practice, you will need to figure out how to make the pieces smaller and give them the right amount of flex and tension.

We hope that new research in micron-scale material manufacturing and work on new materials with adjustable stiffness will lead to advances that will make powerful smart mechanical neural networks with micron-scale elements and dense 3D connections commonplace in the near future.