Oftentimes, college professors are the ones who open doors for their students, recommending them for internships, fellowships and jobs. But a few years ago, Minyi Dai, a PhD student who was working as an intern at Seagate Technology, opened a door for her advisor Jiamian Hu, an associate professor of materials science and engineering at the University of Wisconsin-Madison.
Now, Hu and Seagate researchers have co-authored a paper in the journal npj Computational Materials laying out a computational framework for discovering advanced hard drive materials and processes.
This academic and industry collaboration, Hu says, is a fulfilling addition to his research. “I see my research in two ways. In one way, my work is really fundamental, and we do it just to discover new knowledge with no immediate application,” he says. “The other way is to focus on an application, which is what industry researchers do. It’s been useful to hear what they want to deliver and what research is relevant to them, and if it can contribute to the development of important materials.”
The collaboration began in 2022 after Dai returned to Hu’s lab after a summer internship with Seagate Technology, a leading innovator in mass-capacity data storage. She let Hu know Seagate was looking for new research partners. That led to a visit to Seagate’s California offices, where Hu delivered a presentation.
It turned out that the company was looking for someone with Hu’s expertise in modeling materials microstructures. Seagate and Hu began working together, tackling problems the company has wrestled with for decades.
Like computer chip manufacturers who are always on the hunt for new ways to cram more transistors on microchips to boost their power, disc drive designers are also looking for new materials and methods to maximize the amount of data stored in each millimeter of their disc drives.
One promising method, and one the industry has investigated for more than two decades, is thin films studded with nanoscale pillars of an iron and platinum alloy separated by a nonmagnetic material. Using these specialized materials and a technique called heat-assisted magnetic recording, Seagate has already produced a record-breaking hard drive with a 44-terabyte capacity.
To reach its long-term storage capacity goals, however, the company needs to find processes to better engineer the material, decreasing its thickness and diameter and improve the spacing and regularity of the alloy pillars. The company also needs to find an ideal nonmagnetic material in which to embed its nanopillars. Engineering these structures at the nanoscale, however, has proved difficult because the variables in processing the alloy-studded films are vast. Testing each possibility experimentally would be a long, tedious and expensive process.
That’s where Hu’s work comes in. As an expert in computational materials discovery, he uses high-throughput phase-field simulations to find the right combination of processing variables—like the materials used and the film deposition temperature—to come up with candidate processes that can later be verified experimentally.
The Seagate project, however, was even more complex than usual, requiring many different parameters with many different variations. That meant the simulation, while significantly narrowing the field, still came up with hundreds of thousands of possible processes that could lead to the desired nanostructures.
Many of those candidate processes have issues that make them unworkable, but before discarding them an expert needs to wade through this vast number of possible solutions to assess them individually.
To get around this, Hu added a specialized AI to the system to do that review. “In this new framework, our phase field simulation does the forward predication, and then the AI works backward in a closed-loop workflow,” he says. “If we don’t use AI, we have to look through about 200 processes to produce one solution in a proof-of-concept problem. Using AI, we came up with solutions better and faster than humans, reducing the computing costs tenfold.”
To double check that his method worked, Hu set up a “John Henry” moment: two postdocs—one, an expert in phase field simulation with little experience in AI and the other, an expert in AI who had little experience in phase-field simulation—went head-to- head with the AI to see who could more successfully winnow the results of the simulation.
Hu found that, besides being much faster, the AI outperformed the human researchers, who tossed out several potentially intriguing solutions. “The AI doesn’t have the same biases as the human researchers,” says Hu. “There are things humans might overlook, but the AI suggests there might be more possibilities. That’s a good way to accelerate knowledge and discovery. I’m excited about these findings. With an accurate forward-looking phase field predication and AI, the combined loop can solve real-world problems. That’s why I’m very excited about this work.”
Hu says that working with Seagate has changed the way he views materials science. “It’s been eye-opening to see how industry works,” he says. “In academia, we have more freedom to explore. In our models, we have to make sure every step is scientifically sound, but we have a lot of assumptions in our models, and that might often be okay as long as the core science is captured.”
Industry, however, which must produce products at scale, isn’t focused on doing science for science’s sake. “They are very problem-oriented and have to be really precise,” says Hu. “I admire Seagate’s whole team and the decades of work they’ve done.”
While his method is promising, and he hopes it is useful for Seagate’s internal research in its current state, Hu says it still requires further development. “The model we have is state of the art but it’s not there yet,” he says. “I hope we can do more to push the limits.”
Featured image credit: Seagate Technology
Other UW-Madison authors include Xiangwei Guo, Zebin Li, Mojue Zhang and Jingheng Chai.
Other authors include Wei-Heng Hsu, Timothy Klemmer, Olle Heinonen, Yingguo Peng and Ganping Ju of Seagate Technology.
The authors acknowledge support from Seagate Technology; the National Science Foundation (NSF) under Grant No. DMR-2237884; the NSF-funded Wisconsin MRSEC (DMR-2309000); NSF Grant No. CBET-2412157; the Bridges at the Pittsburgh Supercomputing Center through allocation TG-DMR180076 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, supported by NSF Grants No. 2138259, No. 2138286, No. 2138307, No. 2137603, and No. 2138296.