It’s often said that materials pose the primary bottleneck in the development of new technologies.
This is particularly true for energy, where the advancement of technologies such as batteries, fuel cells, and even nuclear reactors are often dependent on the development of novel materials, making progress beholden to the pace of materials R&D.
However, with the emergence of artificial intelligence, specifically deep learning, and the increasing power of electron microscopy, the pace of innovation has the potential to greatly accelerate.
Consider a team of researchers from the Department of Energy’s Oak Ridge National Laboratory (ORNL) that recently demonstrated the power of deep learning and convolutional neural networks in examining atomically resolved images from an electron microscope and classifying their 2D symmetry at the nanoscale. The work, led by Rama Vasudevan, a materials scientist at the Laboratory’s Center for Nanophase Materials Sciences, was recently featured in the journal npj Computational Materials.
Solid materials exhibit symmetry, a crucial property that allows researchers to reduce the complexity of their description of materials by focusing on a few symmetry modes rather than the 1023 atoms in a solid.
But while the average symmetry class of a material can be routinely determined from x-rays and neutrons, determining its symmetry class at the nanoscale is a challenging computational and experimental problem—yet a necessary one for researchers tackling a range of science and energy challenges.
“Determining the symmetry state of a solid constrains the space of its potential functionalities” said Nouamane Laanait, a former Eugene P. Wigner Fellow and current computational physicist in ORNL’s Computational Sciences and Engineering division who co-authored the study. “By using deep learning, we showed that we can successfully extract the 2D symmetry state of materials at the nanoscale level from electron microscopy without the need of a physics-based model and with minimal user input.”
The research team initially used ORNL’s Compute and Data Environment for Science (CADES) OpenStack Cloud to run their deep learning models. The CADES OpenStack Cloud solution features fully customizable virtual machines (VMs) and enables researchers to leverage self-service portals to rapidly request these VMs for production, testing, and development.
Furthermore, said Laanait, the work serves as a stepping stone toward using deep learning and artificial intelligence to solve challenging inverse problems in the imaging of materials, work that will also require the OpenStack Cloud resources. For example, a team led by Laanait is using deep learning to extract the full 3D symmetry state of materials from electron microscopy/scattering. In collaboration with Christopher Layton, principal system engineer of ORNL’s OpenStack Cloud, the team of researchers used 1,000 cloud CPU cores to prototype different preprocessing routines and physics-based simulation models to generate new training data for nearly 60,000 solid-state materials. Laanait and colleagues are currently training deep learning models on this data using OLCF’s Summit, ORNL’s flagship computing system and the most powerful computer in the world.
“The OpenStack Cloud is a remarkable addition to the computing resources available at ORNL,” said Laanait. “It addresses indispensable research needs more suitable for the cloud than the Laboratory’s high-performance computing resources. The cloud infrastructure enabled by Chris Layton’s efforts is truly outstanding.”
AUTHOR: Scott Jones