Frontiers of Materials That Learn: Proceedings of a Workshop (2026)
Chapter: 5 Looking Forward
5
Looking Forward
Lisa Manning, workshop planning committee chair, moderated the fifth and final session, which had the participants split into breakout sessions for individual discussions, return to the main room to report on what was said in those breakouts, and then hold a discussion period with all of the participants on hand.
For guidance on the breakout discussions, Manning provided a list of three questions that each breakout group was asked to address:
- What are some of the ideas discussed today that resonated with you and would be exciting directions to pursue? Based on your own expertise, are there additional directions or synergies that should be explored?
- What new theoretical, computational, and experimental techniques need to be developed to push this field forward? What is currently limiting progress in this area?
- What new functionalities, applications, and use-cases might emerge from discoveries in this topic area? How might discoveries in this field impact other disciplines?
REPORTS FROM THE BREAKOUT SESSIONS
Each of the three breakout groups had one participant summarize what that group had discussed.
Group 1
The first summary was provided by Itai Cohen of Cornell University. One of the things that his group found most interesting, he said, was the idea of having systems with many parameters. “We were all sort of awestruck by the amount of flexibility that that gives you,” he said.
So they got into a discussion of how to get something useful out of this massive parameter space. How, for instance, does biology emerge from such a space? And although they batted around a lot of ideas, no one really had any answers. For example, Cohen said, no one knew at what granularity constraints should be imposed on the space, although they believed that there must be some set of constraints. One suggestion was that such things as the composition of the atmosphere and the energy budget associated with evolution might “cut through this massive parameter space to give you these lower-dimensional behaviors that we observe in biology,” he said, and that something about those low-dimensional behaviors enabled biological systems to evolve in such a way that they did not have to sample the entire parameter space in order to get something useful.
One fundamental issue that needs to be addressed, Cohen said, is how one should think about the constraint space and what is the granularity. “We were talking about the fact that in computer science people have been exploring this in a particular way, and in other sciences people are trying different approaches,” he said. “Do I need to worry about a computer program at the level of a semicolon which screws up my program, or do I think about it more in terms of the functionality? Do I think about it in terms of an amino acid sequence change, or do I think about it in terms of the fact that this change influenced the behavior of the protein? At what level do we start to worry about the constraints?”
If these issues could be solved, Cohen said, his group believed that it would be possible to develop some powerful applications. The example he gave was of a parallel immune system that would act like a shadow of the natural human immune system, with microscopic robots that could kill dangerous cells and de novo proteins that the body could not recognize and reject but that helped get rid of diseases and improve health.
Group 2
Laurel Kroo, a professor of polymer science and engineering at the University of Massachusetts Amherst, reported on the second group’s discussions. In considering the first question, she said, the group began talking about what learning is, and they decided to think about learning in terms of the ability to explore, “and
we had this kind of deeper dive down into what is nontrivial and then how to find the local rules that lead to those nontrivial behaviors.”
They also had discussions about usefulness and complexity and the requirements for those things. They discussed T cells as a good example of bio learning in a material where the material (i.e., a T cell) is faced with a variety of different problems and can learn and has some degrees of freedom to choose what the problem is based on the environment and the situation in which it finds itself. That in turn led the group to talk about the possibility of, instead of optimizing a material for a particular situation, developing materials with general learning capabilities. Such a material might be able to tackle a class of problems in a particular space, perhaps handling just a few different things or perhaps having a more universal capability.
In addressing the second question of what tools the field needs, the group looked at two different types of tools, experimental tools and then simulation and theory tools. One of the needed experimental tools that the group identified, Kroo said, was molecules other than DNA for molecular-level materials that learn. One of the group’s members, Bartosz Grzybowski of IBS Korea, talked about the “delicacy” of some of these landscapes and how to engineer the ability to explore them; this will be particularly important in chemical systems, “where you can just kind of crash out of solution really easily,” Kroo said, as opposed to biological systems, “where often you can explore for a really long time.” This in turn led them to talk about the importance of high-throughput robotic models of gathering data, she said, which in turn led into a discussion about theory and computational tools.
Next, Kroo said, there was a spirited debate about the usefulness of massive amounts of data and what is necessary to handle all those data in order to get something useful out of them. “What we came to is that there’s this big question of what is the Local Rule in theory,” she said, “and it’s a really hard question to answer.” Eventually they came up with an analogy: “Are you trying to build a map of the phase space for how to get from point A to point B in some evolving network, or are you trying to build a self-driving car that can get there if you tell it just try to find this place?” They used that analogy to delineate between the two different approaches of either mapping enormous multiphase spaces that may be very heterogeneous versus simply zeroing in on a desired answer without needing such a map.
Another thing the group discussed was that learning with physical objects is very different from using traditional artificial intelligence (AI) methods and how the community can sometimes seem somewhat “sheepish” about how they are using different tools in their learning approaches. However, Kroo added, there is a reason that the community is using different tools—it is because they are dealing with physical systems that have conservation laws and various other constraints on the system. Ultimately the community needs to figure out how to formalize its approaches so that it can begin to extrapolate to larger-scale material learning systems.
Concerning the third question on applications, the group came up with a variety of ways to talk about applications of materials that learn in a meaningful way and hopefully inspire funding commitments. They discussed the category of bioinspired functionality of materials, such as self-replicating materials, and also about adaptive solid materials such as artificial bones or prosthetics made with self-healing materials that would heal under various different circumstances. There were also some discussions, Kroo said, about how one could use reconfigurability and adaptability to build weapons or warfare-related materials for different environments. Another category was materials for use in extreme environments, such as molecules that could adapt to an environment and shield a biological system. Finally, one suggestion of a “low-hanging fruit” was biomedical applications such as stents or other items that could immediately help companies in biomedical fields.
Group 3
Doug Durian of the University of Pennsylvania reported for the third group, and for his summary he offered a number of high-level questions that the group discussed and whose answers will likely play a major role in the future of the field.
The first was what should the architecture of learning networks be? It is believed that architecture plays an important role in the performance of such networks, but exactly how does the architecture affect the tasks that a network can learn, and how should networks be combined? “This is a very, very important question, and we don’t have much of a clue,” he said. “We’re just trying things.”
Another important question concerns the role that heterogeneity plays in determining the performance of a network. For example, the brain has many different kinds of neurons, and that heterogeneity, Durian said, plays a big role in “enabling the brain to be this fantastic learning material, much better than anything we can now build in the laboratory.” So how can heterogeneity help with creating materials that are better learners?
The role that stochasticity plays in the performance of a learning material is a similarly important question, he said. “Noise seems to be very important in learning,” Durian said. “We think of the brain as a noisy system, and to me it’s just amazing that the brain can learn in spite of the noise.” But maybe that noise is actually part of the brain’s secret, he continued. Perhaps it is not learning in spite of the noise, but rather it is learning because of the noise. If so, it will be important to figure out how to take advantage of noise in material systems that learn. For example, in artificial networks, stochastic gradient descent is an important tool where noise plays a helpful role. So what other ways might noise help. “That’s a very important thing that we identified is going to drive things forward,” Durian said.
Another important topic the group discussed was timescales and equilibrium, he continued. “The plasticity of the brain and the speed of propagation—those
things are all mingled together.” Physical signals will propagate at certain speeds through a material, influenced by relaxation times, he said, “and you’re adjusting learning degrees of freedom on comparable scales. How do those timescales interact to help or maybe harm it?” So researchers need to consider broad questions of equilibrium, he said.
Another question is what are the broad classes of materials that can learn, and what are the advantages of each. Durian noted that he is used to thinking of networks that learn in terms of the particular types he studies—resistor networks or mechanical networks—but there are also chemical networks that learn as well as other types. “How does this physical substrate, if you will, play into the kinds of things that it can learn?” he asked. Why are brains good at learning some things and resistor networks good at learning others, and what exactly are they good at learning?
Finally, Durian said, the group talked about potential applications of learning materials. What are the killer apps? One particularly valuable application of these materials might be the ability to deal with and adapt to changing conditions. “If the conditions in which things are operating are changing, you need to update,” he said. “You can’t just bake in some rules for input–output relations. That functionality has to change with time.” So, if one can build systems that could learn, they can adapt, and that could be a very valuable property. Of course, he added, it would be a killer app if materials that learn made it possible to do AI with a million times better energy efficiency; that would enable things that cannot be done now. But having qualitatively different applications—not just faster and more energy-efficient, but with qualitatively different abilities, such as being able to adapt to changing conditions—could be a game changer.
DISCUSSION
Cohen offered a couple of thoughts on what Durian had said. First, he said, the brain is even more heterogeneous than Durian said. It is not just that it has different types of neurons; it is that each neuron’s behavior changes depending on the chemicals in its environment. “So you should think of neurons as families of behaviors, just in a single neuron, and that makes it even more complex,” he said.
Concerning the role that noise plays in systems that learn, Cohen said that one possibility is that noise allows a system to not overspecialize. “One thing that these networks are really good at is their universal function approximators, and that means that they’re really good at fitting data,” he said. “In fact, they’re really good at overfitting data, and maybe the noise prevents the overfitting of the data,” which in turn would make it possible to retain the “gross trendline,” that is, the underlying model that governs the data’s behavior.
Viola Vogel of ETH Zürich added that the noise levels go down in the brains of Alzheimer patients and said that some people in the community think that
having a higher noise level is important for suddenly having great ideas. According to this suggestion, getting to a point where an idea rises above the noise level defines creativity somehow, so brains with higher noise levels would have fewer but presumably better ideas.
Next, Manning addressed the group and asked what people thought about where resources should be devoted in the community. Are there some high-risk, high-reward ideas that should be pursued?
One audience member said that she did not think it made sense at this point to devote a lot of resources to specific applications. Instead, there are some fundamental issues that the field needs to address in order to make progress. For instance, how can things be tuned at the local level, particularly if the goal is for the system to self-tune according to a local rule? Implementing such tuning would require, first of all, figuring out a good local rule, given a system and the behavior one wants it to acquire, and then figuring out how to implement that rule. This is highly nontrivial, she said, and by doing this in a number of different systems should allow researchers to start constructing the universal frameworks they need to deal with these systems.
Kroo followed up on that comment by saying that while her group came up with very specific application areas, it was more to demonstrate the power of the potential toolbox and not to say that those specific applications should be funded.
One audience member said that he thought that it was important to have some well-defined objectives and funding because otherwise researchers in the field may be seen as just playing around. “Honestly, I heard about dynamic self-assembly 30 years ago, and we haven’t made concrete progress,” he said. “In the meantime, AI overtook us, and we are still talking about possible applications.” So it is important to achieve some concrete milestones, he said, or else the field may seem to others as little more than philosophizing.
Durian said that having big aspirations, such as building a better brain, is certainly important, but having smaller, more specific goals—“something that can pull us along,” he said—is still an important thing.
Another audience member suggested that one area in which materials that learn might have important applications is exploring space and, particularly, colonizing Mars. A first step will be building a base, she said, and this is a task where it will not be necessary to have super-intelligent agents; lower-intelligence robots that could work collectively should be sufficient.
An audience member, commenting that those in the soft-matter field are “famous for not having big questions,” asked the panel whether such a grand question might arise in exploring how to set the right constraints in a large parameter space in order to define a manifold on which soft-matter systems can exist and have interesting functions. “This is one of these emerging questions that keeps coming up in these types of meetings that I’ve attended,” she said. “Do you think this has
the potential to become one of those grand questions in soft matter, or is it too small for that scale?”
Cohen said he thought this could indeed be such a grand question. “That’s the new stat mech [statistical mechanics], right?” Understanding how a huge number of variables can be contracted down to some sort of behavior is an extremely interesting question, he said.
Manning agreed and said, “As soon as you have all of these massive Avogadro’s number of tunable parameters, it opens up a whole bunch of questions.” For instance, she said, “If you have any connections between those tunable degrees of freedom at the smallest scales and any emergent behavior at all, course graining goes out the window, right?”
Cohen said that a major question is what sorts of structures emerge on these lower-dimensional manifolds. “What are the organizing defects and dynamical trajectories and bifurcations, et cetera, that govern the behaviors on these spaces?” he asked, and he said that it will not be statistical mechanics or critical points that govern the organization, but perhaps it will be bifurcations and related things from dynamical systems theory. Operating near such transitions can be very useful for taking advantage of the sensitivity that exists at these points, he said. “Somehow biology seems to work near these phase transition points and leverages them,” he continued, “and I think that something like that must emerge in these systems as well, and I think that is where the most interesting stat mech questions are. How much can be just imported directly from everything that we’ve done in RG [renormalization groups]? How much needs to be reinvented for these kinds of high-dimensional systems?”
Mentioning Ilya Prigogine’s 1977 Nobel Prize for work in nonequilibrium statistical mechanics and minimal entropy production, Bartosz Grzybowski of IBS Korea said that there has been little meaningful advance in the field since then. So, he asked, “What are our chances of bringing anything new to the table after 50 years of failure?”
“I think we have a good chance,” Cohen said, “because this is the first time that we have algorithms that are powerful enough to take these massive datasets and be able to pull out latent spaces that are low-dimensional and interpretable.” This is a big change from just a decade ago, he added. Then, in response to a followup question about experiments, he said that, in terms of gathering large amounts of data, experimental science is also in a golden age right now. In particular, he named sequencing data and neural recording data as filling up huge datasets. “We are gathering massive amounts of high-dimensional data,” he said, “and the stat mech of our time is to figure out how all of that then leads to these relatively simple behaviors that we can observe in these systems.”
“That is very cool what you are saying,” Grzybowski responded, “because coming from chemistry, I’ll tell you, go to a chemistry lab and the glassware is exactly
what was used 100 years ago, and we are generating massive amounts of chemicals that go down the drain, exactly as 100 years ago.” But chemistry, Cohen replied, is one of the fields that is really poised for a data revolution because of the automation that is being implemented in chemistry labs. “If you can start to do everything automated and explore the phase space, you can imagine using a lot of the information theory techniques that we have to figure out where the phase space is least explored and start doing a much more systematic investigation.”
Arvind Murugan of the University of Chicago drew a distinction between Prigogine’s work and the topic of the workshop. Prigogine looked at structure formation in nonequilibrium systems where there is no purpose. “There’s no objective,” he said. “Just systems do stuff. Sometimes they form structures, sometimes not, and the question is when, and that’s a really interesting question.” But the workshop was more concerned with directed goals, where there is a desired functionality, and the question is how can one nudge a system to acquire that functionality. “I would say this is not ground that’s been well trodden before. So, who knows what we’ll find?”
Manning closed by referring to Murugan’s comment and saying that while the workshop’s stated purpose was to examine the state of the science for materials that learn, it makes sense to look at the issue more broadly in terms of creating materials with a purpose or things that they are trying to do. That is an exciting frontier where more tools are needed in order to keep making progress, she said, and the workshop identified a number of such needed tools.
