Argonne scientists from several disciplines have combined their forces to create a new process for testing and predicting the effects of high temperatures on refractory oxides.
Cast iron melts at around 1,200 degrees Celsius. Stainless steel melts at approximately 1520 degrees Celsius. If you wanted to shape these materials into everyday objects, like the pan in your kitchen or the surgical tools used by doctors, it stands to reason that you would need to create ovens and molds from something that can withstand even at these extreme temperatures.
This is where refractory oxides come in. These ceramic materials can withstand blazing heat and retain their shape, making them useful for everything from furnaces and nuclear reactors to thermal protection tiles in spaceships. . But given the often hazardous environments in which these materials are used, scientists want to understand as much as possible what happens to them at high temperatures, before components constructed from these materials encounter those temperatures in the real world.
A team of researchers from the Argonne National Laboratory of the US Department of Energy (DOE) have devised a way to do this. Using innovative experimental techniques and a new approach to computer simulations, the group has developed a method to not only obtain precise data on the structural changes that these materials undergo near their melting point, but also to predict specifically other changes that currently cannot be measured.
The team’s work has been published in Physical examination letters.
The seed for this collaboration was planted by Marius Stan, leader of the Intelligent Materials Design program in the Applied Materials division of Argonne. Stan’s group had developed many models and simulations on the melting points of refractory oxides, but he wanted to test them.
“It’s rooted in the desire to see whether our mathematical models and simulations represent reality or not,” Stan said. “But that has evolved into a study of machine learning. What I find most exciting is that now there is a way for us to automatically predict interactions between atoms.”
The innovation began by flipping a familiar script, according to Ganesh Sivaraman, lead author of the paper and computer science assistant at Argonne’s Data Science and Learning division. He did this work while a postdoctoral fellow at the Argonne Leadership Computing Facility (ALCF), a user facility of the DOE Office of Science.
While most experiments start with a theoretical model – essentially, an educated, informed estimate of what will happen under real conditions – the team wanted to start this one with experimental data and design their models around that.
Sivaraman tells the story of a famous German mathematician who wanted to learn to swim, so he picked up a book and read about it. Creating theories without considering experimental data, Sivaraman said, is like reading a book about swimming without ever going into a pool. And the Argonne team wanted to get into the deep end.
“It’s more accurate to build a model around experimental data,” Sivaraman said. “It brings the model closer to reality.”
To get this data, computer scientists teamed up with physicist Chris Benmore and assistant physicist Leighanne Gallington of Argonne’s X-ray science division. Benmore and Gallington work at the Advanced Photon Source (APS), a user facility at the DOE Office of Science in Argonne, which generates very bright x-ray beams to illuminate material structures, among other things. The beamline they used for this experiment allows them to examine the local and long-range structure of materials under extreme conditions, such as high temperatures.
Of course, warming the refractory oxides – in this case, hafnium dioxide, which melts at around 2,870 degrees Celsius – has its own complications. Usually the sample would be in a container, but there isn’t one available that would withstand these temperatures while still allowing x-rays to pass through them. And you can’t even put the sample back on a table, because the table will melt before the sample does.
The solution is called aerodynamic levitation and involves scientists using gas to suspend a small spherical sample (2-3mm in diameter) of material about a millimeter in the air.
“We have a nozzle connected to an inert gas stream, and while it suspends the sample, a 400 watt laser heats the material from above,” said Gallington. “You have to tinker with the gas flow to make it levitate stably. You don’t want it to be too low because the sample will hit the nozzle and could melt on it.”
Once the data was taken and the scientists at the beamline fully understood what happens when the hafnium oxide melts, the computer scientists took the balloon and ran with it. Sivaraman fed the data into two sets of machine learning algorithms, one of which understands theory and can make predictions, and another – an active learning algorithm – which acts as an assistant to teaching, giving the first only the most interesting data. work with.
“Active learning helps other types of machine learning learn with less data,” Sivaraman explained. “Let’s say you want to walk from your house to the market. There may be several ways to get there, but you only need to know the shortest route. Active learning will point out the shortest route and filter out the others.”
The calculations were carried out on supercomputers at the ALCF and at the Argonne Computer Resources Laboratory. The team ended up with a computer-generated model based on real data, a model that allows them to predict things that the experimenters didn’t – or couldn’t – capture.
“We have what’s called a multiphase potential, and it can predict a lot of things,” Benmore said. “We can now go ahead and give you other metrics, such as how it retains its shape at high temperatures, which we haven’t measured. We can extrapolate what would happen if we went to the – above the temperature we can reach. “
“The model is as good as the data you give it, and the more you give, the better it gets,” Benmore added. “We are giving as much information as possible, and the model is improving.”
Sivaraman describes this work as a proof of concept, which can nourish other experiences. It is a fine example, he said, of collaboration between different parts of the Argonne and of research that could not be done without the resources of a national laboratory.
“We will repeat this experiment on other materials,” Sivaraman said. “Our APS colleagues have the infrastructure to study how these materials melt under extreme conditions, and we are working with computer scientists to create the streaming software and infrastructure to quickly process these large-scale datasets. . We can integrate active learning into the framework and teach models. to process the data flow more efficiently using ALCF supercomputers. “
For Stan, proof of concept is one that can replace the necessary boredom of people working on these precise calculations. He has seen this technology evolve over the course of his career, and now what used to take months only takes days.
“I’m not saying humans aren’t great,” he chuckled, “but if we get help from computers and software, we can be bigger. This opens the door to other experiments like this that advance science.