ИИ и машинное обучение ускоряют создание материалов с заданными свойствами

When you come to think about it, many of the early ages of human development were defined by the materials people used, such as the Stone Age, the Bronze Age, and the Iron Age. Much later came the Industrial Revolution and the Industrial Age. This was followed by a succession of increasingly overlapping technological eras—the Machine Age, the Atomic Age, the Space Age, the Computer Age, and the Information Age, to name but a few. Even the early ages weren’t as neat and tidy as their names might suggest. They overlapped in all sorts of ways. For example, cultures didn’t switch from one material to another overnight. Instead, they would gradually transition from stone to bronze and then to iron. Also, while one culture continued to bask in its Stone Age, neighboring cultures might already be dipping their toes into their Bronze or Iron Ages. And now, in an ironic twist, we may be coming full circle as we enter what we might call a Materials Age. In this case, however, rather than simply discovering materials and figuring out what we can do with them, we are increasingly able to turn the process on its head. We can start by specifying the characteristics we want—strength, flexibility, conductivity, transparency, heat resistance, magnetism, or some combination thereof—and then set about creating a material that provides them. This may involve combining elements in ways nobody has previously considered, arranging atoms and molecules in novel structures, or creating composites whose constituents work together to provide combinations of properties that would once have seemed mutually exclusive. In effect, instead of asking, “What can we make with this material?” we can increasingly ask, “What material do we need to make this?” I read a lot of science fiction. Neal Stephenson’s The Diamond Age immediately springs to mind because its world has molecular-scale nanotechnology and matter compilers. One of my more recent discoveries is Dennis E. Taylor’s We Are Legion (We Are Bob). Amongst many other things, while our hero Bob no longer has a biological body (bummer), he does have access to futuristic 3D printers. Rather than squirting molten plastic through a nozzle, Bob’s printers can create objects one atom at a time, assuming they have access to the necessary raw materials and energy. I can easily envisage a future in which we are able to create materials in much the same way, atom by atom, on an industrial scale. When that time comes, our descendants may look back on today’s state-of-the-art materials science in much the same way we might regard tossing a bunch of ingredients into a barrel and stirring them with a big stick (have you ever had the opportunity to sample my mother-in-law’s cooking?). Of course, things aren’t really that bad (I’m talking about the current state of affairs in materials science, not my mother-in-law’s cooking). At the moment, aided by increasingly sophisticated modeling and simulation, we largely create new materials by combining promising ingredients, experimenting with different structures and manufacturing processes, observing which properties emerge, and seeing if we can repeat the process. But this is starting to change… “Really?” you cry. “Tell me more!” Well, I actually had other plans, but I don’t wish to be churlish, so if you insist, I shall share some recent intelligence of which I was just made aware. The source of this aforementioned intelligence was Dr. Joshua Young, Senior Application Scientist at Matlantis. Josh is a fun guy to chat with. He kicked off by saying, “You can just call me Josh.” I responded, “You can just call me Max the Magnificent.” We were off to a good start. Josh explained that the company’s name is a play on Atlantis because, as Josh put it, “we’re searching for the lost materials.” They’ve even got a handy-dandy video that provides a high-level view of what the company does. As part of his role, Josh helps researchers turn real-world materials questions and problems into things that can be investigated using atomistic simulation. In the simplest terms, this means creating a computer model in which every atom in the system is explicitly represented. Don’t panic. We aren’t talking about modeling every proton, neutron, and electron individually. If our system contained a water molecule, for example, we would explicitly represent its two hydrogen atoms and one oxygen atom, along with the ways in which they interact. One well-established way of modeling these interactions is Density Functional Theory (DFT), a quantum-mechanical technique that works with the distribution—or density—of electrons throughout a system. DFT can calculate the energies and forces associated with the atoms, along with electronic properties such as conductivity and, in the case of semiconductors, the all-important band gap. The downside is that DFT is computationally expensive. In broad terms, the amount of work involved scales approximately with the cube of the number of atoms being modeled. Increasing the size of a system by a factor of ten may increase the computational effort by a factor of 1,000 or more. Eeek! And this is where Machine Learning Interatomic Potentials (MLIPs) strut onto center stage with a fanfare of flügelhorns (the cheeky little rascals). In a crunchy nutshell, the idea is to perform humongous amounts of computationally expensive DFT calculations on carefully selected atomic structures, then use the results to train a machine-learning model to recognize the relationships between the chemistry of a system, the way its atoms are arranged, and the resulting energies and forces. Once trained, the MLIP can be presented with a structure it hasn’t previously encountered and infer the energies and forces that a DFT calculation would be expected to produce. Better yet, while DFT scales approximately cubically with the number of atoms, MLIPs scale approximately linearly, allowing researchers to simulate much larger and more complex systems. There is a catch. Although an MLIP can predict energies and forces, it doesn’t retain the detailed electron-density information provided by DFT. Thus, it cannot directly calculate electronic properties such as a semiconductor’s band gap. Since electronic properties are rather important when your day job involves creating electronic devices, it’s easy to see why some semiconductor researchers may have regarded MLIPs as interesting but of limited use (“close, but no cigar,” as it were). The change in thinking is that this doesn’t have to be an either/or decision. Suppose we want to investigate defects in a semiconductor crystal. An MLIP can rapidly explore where the defects might occur and how the surrounding atoms will rearrange themselves. Once we’ve identified a handful of interesting possibilities, DFT can be brought back into play to determine their detailed electronic properties. All of which brings us back to the guys and gals at Matlantis. At the heart of the Matlantis platform is an MLIP called PFP, which stands for Preferred Potential. This is where things start to get seriously impressive because PFP was trained using approximately 62 million DFT calculations spanning an enormous range of chemical compositions and atomic structures. Even 62 million calculations represent only a teensy-weensy fraction of all the ways in which the 96 elements under consideration (we can ignore the ones that disappear almost as soon as you make them) could conceivably be combined and arranged. Fortunately, you don’t need to calculate every possibility. Josh explained that there are two dimensions to consider: chemical space (what elements are present) and structural space (how the atoms are arranged). If you intelligently choose enough representative examples from both, the machine-learning model can make predictions about structures it hasn’t previously seen. This broad training is why the folks at Matlantis describe PFP as “universal.” Rather than being designed solely for semiconductors, batteries, catalysts, or any other particular class of materials, it can be applied across a wide range of problems. Semiconductor researchers can investigate deposition, etching, defects, and interfaces. Battery developers can study how ions move through materials. Chemical companies can investigate catalysts and reactions. Other applications include metals and alloys, coatings, membranes, pharmaceuticals, polymers, and liquids. The common denominator is that, somewhere at the bottom of the problem pile, atoms are interacting with one another. Researcher operation screen (Source: Matlantis) Matlantis itself is a cloud-based platform, so users don’t need to assemble their own GPU farms before they can start doing useful work. Furthermore, the folks at Matlantis don’t simply toss the software over the fence and run for the hills. Josh described their approach as “high-touch customer support,” helping researchers translate their real-world problems into useful simulations. Alternatively, Matlantis can undertake contract research. A customer might say, “We have millions of possible materials; which ones should we actually make?” Or researchers may have observed something unexpected in the laboratory and come to the folks at Matlantis asking, “Why did this happen?” This is where things start to tie back to our Materials Age. Rather than synthesizing and testing thousands or millions of candidates, researchers can use AI-powered simulations to whittle an enormous search space down to a handful of promising possibilities. Instead of saying, “Here are ten million materials; let’s start making them and see what happens,” we can increasingly say, “Here are ten million possibilities. Tell me which ten are worth making.” And things are getting even more interesting as agentic AI joins the party. Matlantis now allows researchers to use AI coding agents such as Claude and Codex inside the platform. The company provides the AI “skills” that teach these agents how to work with Matlantis, allowing researchers to describe what they want to accomplish using natural language. As an off-the-wall example, imagine telling an AI, “I need a material that is extremely strong, weighs almost nothing, remains stable at 1,000°C, and is transparent to visible light.” The AI could explore huge regions of chemical and structural space. MLIPs could rapidly eliminate unpromising candidates. DFT could examine the survivors in detail. And the most promising candidates could then be synthesized and tested in an automated laboratory, with the results feeding back into the next round of simulations. This isn’t the same thing as Bob’s futuristic 3D printers creating arbitrary objects one atom at a time. We aren’t there yet—not by a long chalk—but may I make so bold as to suggest that the folks at Matlantis are taking some of the first steps along the path that might eventually take us there?