New machine-learning algorithm goes against the grain

Modern scientific research on materials relies heavily on exploring their behavior at the atomic and molecular scales. For that reason, scientists are constantly on the hunt for new and improved methods for data gathering and analysis of materials at those scales.

Researchers at the Center for Nanoscale Materials (CNM), a US Department of Energy (DOE) Office of Science User Facility located at the DOE's Argonne National Laboratory, have now developed a machine-learning algorithm for quantitatively characterizing, in three dimensions, materials with features as small as nanometers. Researchers can apply this pivotal discovery to the analysis of most structural materials of interest to industry.

"What makes our algorithm unique is that if you start with a material for which you know essentially nothing about the microstructure, it will, within seconds, tell the user the exact microstructure in all three dimensions," said Subramanian Sankaranarayanan, leader of the CNM theory and modeling group and an associate professor in the Department of Mechanical and Industrial Engineering at the University of Illinois at Chicago.

"For example, with data analyzed by our 3D tool users can detect faults and cracks and potentially predict the lifetimes under different stresses and strains for all kinds of structural materials," said Henry Chan, CNM postdoctoral researcher and lead author of a paper on the algorithm in npj Computational Materials.

Most structural materials are polycrystalline, made up of millions of separate crystalline grains. The size and distribution of the grains and voids in a sample are critical microstructural features that affect important physical, mechanical, optical, chemical and thermal properties. Such knowledge is important, for example, to the discovery of new materials with desired properties, such as stronger and harder machine components that last longer.

In the past, scientists have visualized 3D microstructural features within a material by taking snapshots at the microscale of many 2D slices through the material, processing the individual slices, and then pasting them together to form a 3D picture. That process, however, is inefficient and leads to the loss of information. Researchers have thus been searching for better methods for 3D analyses.

"At first," said Mathew Cherukara, an assistant scientist at CNM, "we thought of designing an intercept-based algorithm to search for all the boundaries among the numerous grains in the sample until mapping the entire microstructure in all three dimensions, but as you can imagine, with millions of grains that is extraordinarily time-consuming and inefficient."

"The beauty of our machine-learning algorithm is that it uses an unsupervised algorithm to handle the boundary problem and produce highly accurate results with high efficiency," explained Chan. "Coupled with down-sampling techniques, it only takes seconds to process large 3D samples and obtain precise microstructural information that is robust and resilient to noise."

The team successfully tested the algorithm by comparing it against data obtained from the analyses of several different metals (aluminum, iron, silicon and titanium) and soft materials (polymers and micelles). These data came from earlier published experiments, as well as from computer simulations run at two DOE Office of Science User Facilities: the Argonne Leadership Computing Facility and the National Energy Research Scientific Computing Center. Also used in this research were the Laboratory Computing Resource Center at Argonne and the Carbon Cluster in CNM.

"For researchers using our tool, the main advantage is not just the impressive 3D image generated but, more importantly, the detailed characterization data," said Sankaranarayanan. "They can even quantitatively and visually track the evolution of a microstructure as it changes in real time."

The machine-learning algorithm is not restricted to solids. The team has extended it to include characterization of the distribution of molecular clusters in fluids with important energy, chemical and biological applications.

This machine-learning tool should prove especially impactful for future real-time analysis of data obtained from large materials characterization facilities, such as the Advanced Photon Source, another DOE Office of Science User Facility at Argonne, and other synchrotrons around the world.

This story is adapted from material from Argonne National Laboratory, with editorial changes made by Materials Today. The views expressed in this article do not necessarily represent those of Elsevier. Link to original source.