For millions of years, living organisms have developed an impressive ability to change their shape and behavior when circumstances change. Cells can alter their structure, tissues can respond to forces, and animals can adapt their movements to their surroundings. This ability to adjust is an important part of survival and evolution.
Now, researchers are bringing a similar idea into the world of artificial materials.
A team led by Yao Du has developed metamaterials that can physically learn how to change shape. Unlike conventional materials, which are normally designed to perform a specific movement or function, these new metamaterials can be trained using examples. They can gradually modify their internal properties, remember learned behaviors, forget old ones, and learn new shape-changing responses.
The research could open a new direction for adaptive materials, soft robotics and physical artificial intelligence.
What Makes These Metamaterials Different?
Metamaterials are specially engineered structures whose unusual properties come from their internal architecture rather than simply the material they are made from.
For example, instead of creating a solid block of material, scientists can build a carefully designed network of small structures. Changing the geometry or mechanical properties of these structures can control how the entire material responds when a force is applied.
Traditionally, researchers decide these properties during the design stage. Once manufactured, the material generally behaves according to that original design.
The new approach changes this idea.
Instead of telling the material exactly how it should behave from the beginning, the researchers allow it to learn the desired shape-changing behavior from examples.
In simple terms, the material is given a target response and gradually adjusts itself until it can reproduce that response.
The Material Learns Through Its Internal Stiffness
One of the most important parts of the system is something called local stiffness.
Think of the metamaterial as being made from many interconnected regions. Each region can be relatively stiff or flexible. These local differences affect how the entire structure bends, stretches and changes shape.
The researchers use these local stiffnesses as the material's internal learning degrees of freedom.
During training, the material progressively updates these stiffness values.
This means the learning does not have to happen entirely inside a traditional computer. Instead, part of the learning process is physically built into the material itself.
That is a major conceptual difference.
Rather than having a robot calculate every movement and then command its material to move, the material itself can store information about how it should respond.
Learning From Examples
The researchers use a technique known as contrastive learning.
The basic idea is to show the metamaterial examples of desired shape changes and use differences between responses to guide how its internal properties should change.
Imagine teaching a flexible structure to bend into a particular shape.
Initially, its response may be far from the target. The internal stiffness distribution is then adjusted. After repeated learning steps, the structure becomes increasingly capable of producing the desired transformation.
This is somewhat similar to training a machine-learning system, but here the information is encoded into the physical properties of the material itself.
The result is a material that doesn't simply follow a fixed mechanical program. It can acquire new physical responses.
The Material Can Forget and Learn Again
Perhaps one of the most interesting features is that the metamaterial is not limited to learning only once.
The researchers demonstrate that it can forget previously learned shape changes and learn new ones in sequence.
This is important because real-world environments are constantly changing.
A conventional structure designed for one particular movement may not easily adapt if its job changes. A physically learning material, however, could potentially be retrained for a different response.
For example, a robotic component could initially be trained to grip one type of object and later be reconfigured to respond differently.
The ability to learn, forget and relearn brings artificial materials closer to the flexibility seen in biological systems.
Breaking the Rules of Reciprocal Motion
The researchers also demonstrate another unusual capability: the metamaterials can learn shape changes that break reciprocity.
In ordinary reciprocal behavior, reversing a process tends to produce a corresponding reverse motion.
But some useful movements in nature are non-reciprocal. Walking is a simple example. A creature can move forward through a sequence of movements even though simply reversing those movements would not produce the same useful motion.
Learning non-reciprocal shape changes could therefore be valuable for artificial systems that need to move through their environment.
Instead of relying only on conventional motors and complicated control systems, future materials could potentially generate useful directional movement through their learned mechanical responses.
Learning Multiple Stable Shapes
The metamaterials can also learn multistable shape changes.
A multistable structure can have more than one stable configuration. Think of a pop-up object or a snap-through mechanism that can remain in different positions without continuously applying force.
When combined with physical learning, this becomes particularly interesting.
The material can learn how to transition between different stable states and use those transitions to perform useful actions.
This gives researchers another way to create mechanical systems that can store and execute complex behaviors without requiring conventional electronic control for every individual movement.
From Learning Material to Robot-Like Actions
The researchers demonstrate that these capabilities can be used for practical mechanical behaviors, including reflex gripping and locomotion.
A reflex grip is a rapid mechanical response where a structure reacts to an object or force and changes its shape to hold it.
In a conventional robot, such behavior could require sensors, processors, software and motors.
A physically learning metamaterial offers a different possibility.
The material itself could potentially encode part of the response. When the appropriate force or mechanical condition appears, the structure could automatically move toward a learned configuration.
The same concept can be applied to locomotion.
By learning appropriate sequences of shape changes, metamaterials could potentially produce movement without relying entirely on traditional robotic control architectures.
A New Way to Think About Smart Materials
The significance of this research goes beyond creating a material that bends in interesting ways.
For decades, smart materials have generally been designed around predetermined functions. Scientists calculate how a material should behave and then manufacture it accordingly.
Physical learning introduces another possibility:
What if materials could be trained instead of completely programmed?
That could make future materials more adaptable.
Instead of manufacturing a different structure for every application, researchers could potentially create a general-purpose learning metamaterial and train it for different tasks.
This could be particularly useful in soft robotics, wearable devices, adaptive structures and machines that operate in unpredictable environments.
Toward Physically Intelligent Robots
The research establishes metamaterials as a promising platform for physical learning.
The long-term vision could involve robots where intelligence is distributed throughout the machine rather than concentrated entirely inside a computer.
A material could sense mechanical conditions through its structure, modify its response and retain learned behavior. This could reduce the complexity of robotic systems and potentially make them more responsive to changing environments.
Of course, significant challenges remain before such materials become common in real-world robots. Researchers will need to explore how reliably these systems can learn, how many different behaviors they can store, how quickly they can adapt and how well they perform outside controlled laboratory conditions.
Nevertheless, the underlying idea is remarkable.
Instead of designing a material once and expecting it to behave the same way forever, scientists are beginning to create materials that can learn how they should behave.
That shift—from fixed mechanical design to adaptive physical learning—could become an important step toward a new generation of intelligent materials and robots.
Reference: Du, Y., van Mastrigt, R., Veenstra, J. et al. Metamaterials that learn to change shape. Nat. Phys. 22, 784–790 (2026). https://doi.org/10.1038/s41567-026-03226-2
Comments
Post a Comment