May 5, 2021 — At the Department of Energy’s Oak Ridge National Laboratory, scientists use artificial intelligence, or AI, to accelerate the discovery and development of materials for energy and information technologies.
“AI gives scientists the ability to extract insights from an ever-expanding volume of data,” said David Womble, ORNL’s AI program director. “New AI tools, together with world-class computing capabilities, are critical to maintaining scientific leadership.”
AI uses computers to mine mountains of data for scientific and engineering insights. Starting with high-quality data matters. Well-characterized materials create a strong knowledge foundation for the design of new materials that launch technologies and expand economies. ORNL has a history of materials development dating back to World War II and a rich archive of data generated on world-class instruments by expert researchers. Increasingly, researchers generate high-resolution materials data at a volume, variety and velocity they never before have had to tackle.
“Ten years ago, a Ph.D. student working on steels might analyze five precipitates a day,” said electron microscopist Chad Parish of ORNL. Such precipitates could embrittle an alloy and cause it to fail. “Now we’ve developed a technique that lets us do a thousand precipitates in five hours. We’re drowning in data. AI may hold the key to making the best use of it all.”
Two types of AI help make sense of big data. Machine learning runs algorithms on high-performance computers to find correlations within large data sets and determine how well they match expectations. In doing so, it reveals features that traditional data analyses may miss because they are subtle, infrequent, complex or unexpected. A step further, deep learning models the workings of the human brain (e.g., applying logic and expertise) to distinguish features in data sets that improve discovery, learning and decision making.
“We can now design machines to do the work that once required a human expert, except much faster and on a larger scale,” said ORNL materials scientist Stephen Jesse.
ORNL researchers have stood at the forefront of efforts to harness machines to propel progress in materials science. Starting in 1992, Bobby Sumpter worked on foundational theory and chemical/materials science aspects of machine learning. Markus Eisenbach joined him in creating the machine learning basis for integrating imaging instruments and high-performance computers. They ran theory-based models on supercomputers and validated the results against experimental findings.
In 2001, when the Materials Research Society issued a conference proceeding on AI methods in materials science, ORNL researchers were well represented, advancing methods to analyze, compress and visualize multidimensional data.
At ORNL’s Center for Nanophase Materials Sciences, Sergei Kalinin, a founding member of the American Physical Society topical group on data science, works with colleagues to pioneer automated analysis of growing data from high-resolution microscopy experiments. “We turned to machine learning methods because traditional approaches were not practical or sufficient,” Kalinin said.
Around 2008, ORNL researchers began publishing papers advancing machine learning and deep learning in processing big data from microscopy and tying experimental results to theoretical models. This effort grew over the subsequent decade to include AI advances such as:
- Complex scanning probe microscope imaging and spectroscopy methods to reveal nanoscale properties in greater detail
- Complete capture of big data streams from microscope detectors
- Workflows for on-the-fly analytics of scanning transmission electron microscopy data
- Automated conversion of microscopy data into libraries of structures and defects
- Algorithms for learning physical laws from observational data
- Assistance to tune microscopes, choose regions of interest in samples and control atom-by-atom assembly
“We are still just scratching the surface with the use of deep learning for quantitative structural analysis of microscopy data,” ORNL’s Albina Borisevich said. “If we can transition from isolated problems to a more general approach, it can completely revolutionize the field.”
For example, ORNL researchers Wei-Ren Chen and Changwoo Do at the Spallation Neutron Source use machine learning to assist in small-angle neutron scattering characterization of a wide range of material structures. The machine learning methods may help them suggest models for data analysis.
ORNL researchers such as Suhas Somnath also have investigated ways to share data widely. He scales codes to run on distributed computing architectures and develops data infrastructure solutions.
“Continual advancements in automation, computational power, and resolution and speed of detectors in instruments now result in ever larger, numerous, more diverse and complex data sets from both simulations and experiments,” said Somnath. “DataFed and the CADES Data Gateway will imminently facilitate collaborative collection, curation, annotation and sharing of data.”
The Summit supercomputer at the Oak Ridge Leadership Computing Facility is ideally suited for training and deployment of AI algorithms on large data sets owing to its 27,648 state-of-art graphics processing units, high-speed file system and large memory. A recent materials microscopy application demonstrated AI scaled to use all of Summit while running at 93% efficiency.
Quality in, quality out
“The major focus in AI tends to be on data analytics, but we should emphasize that the data itself is important,” said ORNL materials scientist Dongwon Shin, who runs thermodynamic models on supercomputers to design high-performance alloys.
He said the ORNL advantage is akin to “grandma knowledge.” You may follow a cookie recipe to the letter, but your grandmother — with her in-depth knowledge of ingredient interactions, etc. — will out-bake you every time. Likewise, ORNL researchers who have worked on materials for decades have world-class data sets with detailed pedigrees.
Shin realized that most machine learning tools were developed by and for programming experts, not the domain scientists. His team developed an open-source toolkit called ASCENDS that lets scientists with little knowledge of programming or data science apply data analytics as easily as using Excel. ASCENDS analyzes correlations between input features and target properties to facilitate the generation and validation of hypotheses and training of machine learning models that predict materials behavior.
Visualizing material success
Visualizing big data is an additional challenge. Materials scientists often use software that comes with the instruments they buy. “Much of the vendor software presents the data collected by instruments in a bad way,” said ORNL’s Philip Edmondson, who investigates materials for nuclear fission and fusion applications.
The scientific community is clamoring for open-source software to help turn big data into something the human mind can interpret. Edmondson and Parish have recommended best practices for improving data visualization.
Materials for advanced nuclear reactors are irradiated in ORNL’s High Flux Isotope Reactor. Then scientists characterize the specimens in detail, and machine learning methods analyze the measurements to determine how irradiation changes the microstructures and properties that are likely to affect the lifetimes of fission or fusion energy systems. “With nuclear materials, there might be millions of dollars and five or more years of investment behind getting one three-millimeter sample into the electron microscope,” Parish explained. “You want to make sure that you’re gleaning all of the scientific insight you can from that sample.”
“We’re investing a lot of money and time into collecting good data,” Edmondson said. “Let’s understand it.”
UT-Battelle manages ORNL for the Department of Energy’s Office of Science, the single largest supporter of basic research in the physical sciences in the United States. The Office of Science is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science.