Frontiers of Materials That Learn: Proceedings of a Workshop (2026)
Chapter: 1 Introduction
1
Introduction
On October 2, 2025, the National Academies of Sciences, Engineering, and Medicine held a hybrid workshop in Washington, DC, and online called Frontiers of Materials That Learn.1 As the workshop planning committee chair, Lisa Manning of Syracuse University, explained in her opening remarks, the workshop was held in response to two trends that have appeared over the past few years. First, she said, a large group of researchers has been thinking about and working with physical learning, which she defined as “onboard autonomous changes to local properties of the system in order to learn a task.” Researchers have studied examples of physical learning in a variety of systems, including resistor networks, DNA self-assembly, origami, and elastic bar networks.
At the same time, she continued, researchers have created types of materials that have the capacity to sense and tune internal parameters autonomously, which is, she said, “exactly what you need in order to be able to do physical learning.” However, there have not yet been many conversations between the two groups of researchers, those studying physical learning and those developing materials of the type that could be useful in developing material learning systems. The goal of the workshop, she said, was to bring these two groups of researchers together in the same room and “see what emerges.”
The National Academies’ Condensed Matter and Materials Research Committee (CMMRC), which Manning also co-chairs, develops annual workshop ideas
___________________
1 See the workshop event page for more information: https://www.nationalacademies.org/projects/DEPS-BPA-24-02/event/45619.
with the goal, Manning explained, to focus on emerging topics that the community is just starting to think about. The topics are identified via multiple rounds of discussions between CMMRC members and stakeholders thus involving many people from many different backgrounds. The topic of materials that learn was one “that a lot of folks were excited about,” she said.
To meet the workshop’s statement of task (Box 1-1), the planning committee assembled a collection of experts in multiple fields that touch on the topic of materials that learn, including physics, chemistry, bio and molecular engineering, mechanical engineering, and computer science and neural networks. A major task of the workshop, Manning said, was to speculate on future directions for the field, including both exciting new areas and any roadblocks that need to be addressed to make progress in this area.
MATERIALS THAT LEARN
To put the workshop’s presentations and discussions into a broad context, planning committee member Andrea Liu gave a brief overview of materials that learn and what is required to create such materials. As part of that overview, she emphasized that a key property of materials that learn is that they need to be tunable. Explaining what it means for a network to be tunable, she said that the idea of a tunable network goes back to the Hopfield networks created by John Hopfield in the early 1980s. “If you imagine an Ising model,” Liu said, “you have spins—physical degrees of freedom—that can point up or down, but the key part is that you have these tunable couplings … which you can adjust.” Those tunable couplings are at the heart of a tunable network, she said, and in Hopfield’s case the couplings were adjusted in a way that allowed the system to store memories.
“That step of making the couplings individually adjustable, instead of being fixed, changes everything,” Liu said. In the usual Ising model, or spin glass, where the interactions between the spins are fixed, a system with 1,000 spins has
BOX 1-1
Statement of Task
An ad hoc planning committee of the National Academies of Sciences, Engineering, and Medicine will organize and conduct a public workshop that presents a high-level synopsis for scientists and policy makers of recent research into physical learning in materials (e.g., the integration of sensing and adaptation in situ to create materials that respond to new, never-seen environments or perform computations). The planning committee will develop the agenda for the workshop, select and invite speakers and participants, and moderate the discussions. A designated rapporteur will prepare workshop proceedings.
essentially the same characteristics as a system with 1 million spins or one with 1020 spins. A Hopfield network with tunable couplings is very different, however. A system of 1,000 spins can store on the order of 1,000 memories, while a system of 1 million spins can store on the order of 1 million memories. “The bigger it gets, the more memories it can store, and that’s really very different,” Liu said. “Its properties are changing. Its behavior is getting more complex as it’s getting bigger.”
This feature is also a property of brains, which are what inspired Hopfield, Liu continued. If one moves from Caenorhabditis elegans (also known as a roundworm), which has 302 neurons, to humans, whose brains have approximately 86 billion neurons, there is a huge difference. Similarly, Liu said, compared with the neural networks of the 1990s, which were relatively tiny, today’s much larger neural networks are far more powerful and can do much more complex things. “So that’s a big difference that comes from having these tunable degrees of freedom,” she said.
Liu next offered a brief explanation of how neural networks work. “They really are solving inverse input–output problems,” she said. For instance, one can input a large set of numbers that represent the pixels of an image, and the output of the neural network will say what the image is. A classic example is a simple neural network that has been taught to tell the difference between an image of a dog and an image of a cat.
How does a neural network do this? Liu explained that the network has a number of tunable degrees of freedom, which are the node weights and biases, and these are tuned iteratively through a learning process until the network produces correct answers. That is, in the beginning a neural network will have no idea of the difference between dogs and cats, and it will answer randomly, but with each answer the tunable parameters are adjusted to encourage the behavior that produced a correct answer and discourage the behavior that led to an incorrect answer. This is done, Liu explained, with the use of a cost function, which is the square of the difference between the system’s actual output and the desired output; the goal is to minimize the cost function—get it as close to zero as possible—so that the system’s output gets to the point where it is providing the correct answer every time, or as close to that as possible. Minimizing the cost function involves an optimization process, she said, “and the bigger the system gets, the more tunable degrees of freedom there are, the more constraints you can satisfy, so the more complex the behavior can get.”
Liu commented that if all parameters in the system were the same or followed some periodic pattern, the network would not provide the correct answer, and if the parameters were fully disordered, the network would not provide the correct answer. “You need correlations in the disorder to get this,” she said.
One advantage of tunable matter, Liu continued, is that it provides a unifying framework for understanding biological function. However, the question that the workshop was designed to address is how to design materials that are like biological materials in their complexity and that can learn and sense and respond to their
environment. The basic approach to doing that, she said, is to look for systems with many tunable degrees of freedom that characterize the interactions and then figure out how those parameters should be tuned to get the desired function. One could do those calculations with an external processor, she said, but ideally one wants a system that will tune itself. This kind of closed-loop tuning is one of the things the workshop would address, she added, noting that there are many different possible tuning processes.
In the case of biology, evolution serves as such a tuning process. “There are also ways to tune using local rules, which, of course, is how the brain learns,” Liu said. “For example, the Hebbian rule that ‘Neurons that fire together, wire together’ is a local rule. You don’t have to know everything about the rest of the system in order to know how to tune yourself. So we’ll hear about those.”
Finally, Liu said, the workshop would address the ingredients that are necessary for this kind of learning. A material system must have reconfigurable, tunable degrees of freedom, for instance. It must also have an ability to sense and respond to the environment. “So this is a huge challenge,” she said. “We don’t have the theoretical, computational, and experimental tools to confront this kind of matter, and it’s a fascinating time.”
ORGANIZATION OF THE PROCEEDINGS
This proceedings follows the structure of the workshop, which consisted of three sessions with speakers’ presentations plus a discussion period as well as one additional session in which the workshop attendees split into three breakout groups to address questions about the future of the field and then reconvened to report on what each group had discussed, followed by a discussion period.2 Thus, Chapter 2 recounts the first session, which looked at physical learning implemented in physical systems. Chapter 3 covers the second session, which was on biological materials as substrates for intelligent behavior. Chapter 4 summarizes the third session, on physical and chemical systems as substrates for intelligent behavior. Chapter 5 reports on the final session with its breakout groups and workshop-ending discussion.
___________________
2 This proceedings was prepared by the workshop rapporteur as a factual summary of what occurred at the workshop. The planning committee’s role was limited to planning and convening the workshop. The views contained in the proceedings are those of individual workshop participants and do not necessarily represent the views of all workshop participants, the planning committee, or the National Academies of Sciences, Engineering, and Medicine.
