Photo credit: www.sciencedaily.com
Researchers at Lehigh University have made significant strides in predicting abnormal grain growth in simulated polycrystalline materials. This groundbreaking achievement could lead to the development of enhanced materials for high-stress applications, such as those found in combustion engines. Their findings were detailed in a recent publication in Nature Computational Materials.
“Our simulations not only allowed us to predict abnormal grain growth but did so significantly ahead of its occurrence,” explained Brian Y. Chen, associate professor of computer science and engineering at Lehigh’s P.C. Rossin College of Engineering and Applied Science and a co-author of the study. “In 86 percent of our observations, we could ascertain within the first 20 percent of a material’s lifespan whether a grain would become abnormal.”
Under continuous heat conditions—such as those reached in rocket or airplane engines—metals and ceramics can fail. These materials consist of crystals or grains that can change in response to heat, with atoms moving in ways that cause the crystals to expand or contract. Abnormal growth in a few grains can significantly affect the material’s properties, potentially reducing flexibility and increasing brittleness.
“Our aim in material design is to intentionally avoid abnormal grain growth,” Chen noted.
A smarter way to identify stable materials
However, the prediction of abnormal grain growth has historically been likened to finding a needle in a haystack. The vast array of potential alloy compositions requires extensive testing that is often impractical and costly. The computational model developed by Chen’s team streamlines this process by quickly discarding materials likely to experience abnormal grain growth.
“The importance of our results lies in efficiently navigating through the extensive array of materials,” he remarked. “We aim to minimize simulation time and swiftly advance our selection process.”
The inherent rarity of abnormal grain growth complicates prediction, as early-stage grains mimicking normal growth patterns can ultimately lead to failure.
Unlocking hidden patterns with AI
To tackle this challenge, the research team employed a deep learning model that integrated two analytical techniques to examine how grains evolve and interact over time. A long short-term memory (LSTM) network was utilized to assess the properties of the materials, while a graph-based convolutional network (GCRN) identified interrelations within the data for predictive purposes.
Initially, the researchers’ goal was merely to achieve successful predictions. The prospect of making early predictions exceeded their expectations.
“We were concerned the data might convey too much noise,” he reflected. “We weren’t sure if the features we investigated would hint at future abnormalities or if such events would be too subtle until they were imminent. However, we were pleased to find that we could make predictions far in advance.”
Effective early detection hinged on the team’s ability to analyze grain characteristics over time prior to the emergence of abnormalities.
“A more useful perspective on grains becoming abnormal is to consider their pre-transition evolution,” Chen explained. “For example, by examining them 10 million time steps before an abnormality, we could detect properties that indicated deviation from those at earlier stages, such as 40 million time steps.”
The researchers synchronized their simulations to the precise moment of abnormal grain growth, analyzing the evolving properties leading up to that point. By uncovering consistent trends in these characteristics, they successfully predicted which grains would become problematic.
“When evaluating grains based on the time leading up to their transition, we discovered shared patterns that proved valuable for predictions,” he added.
In this study, Chen’s team simulated realistic material conditions. The next objective is to apply this methodology to actual material images to confirm their predictive accuracy. Ultimately, Chen aims to identify stable materials that retain physical integrity under extreme high-temperature and high-stress settings. Such advancements could enhance the performance and longevity of engine components.
Furthermore, the researchers recognize the broader implications of their machine learning approach, which could extend beyond materials science to predict other rare phenomena. This methodology might aid in forecasting changes in material phases, evolutionary mutations leading to pathogens, or unexpected shifts in atmospheric conditions.
“This research signifies a groundbreaking opportunity for material scientists to ‘see into the future’ and predict the evolution of material structures in ways previously unattainable,” remarked Martin Harmer, Emeritus Alcoa Foundation Professor of Materials Science and Engineering and co-author of the study. “Its implications are vast, influencing the design of dependable materials across defense, aerospace, and commercial sectors.”
Source
www.sciencedaily.com