According to Joshua Agar, a faculty member of the Faculty of Materials Science and Engineering at Lehigh University, understanding the relationship between structure and properties is an important goal of materials research. However, due to the complexity and multidimensional nature of the structure, there are currently no metrics to understand the structure of the material.
Artificial neural networks, a type of machine learning, can be trained to identify similarities and correlate parameters such as structure and properties, but there are two major challenges, Agar says. One is that most of the vast amount of data generated by material experiments is not analyzed. This is primarily because such images created by scientists in laboratories around the world are rarely stored in a usable way and are usually not shared with other research teams. The second challenge is neural network It is not very effective in learning symmetry and periodicity (how periodic the structure of the material is). Materials researcher..
Today, a team led by Lehigh University has developed a new machine learning approach that allows the prediction of similarity through machine learning. This allows researchers to search unstructured image databases for the first time and identify trends. Agar and his collaborators developed and trained a neural network model with symmetry-aware features, and 25,133 piezoelectric response microscopic images collected on various material systems over a five-year period at the University of California, Berkeley. I applied that method to a set of. Results: We were able to group similar classes of materials, observe trends, and form the basis for beginning to understand the relationship between structure and properties.
“One of the novelties of our work is to build a special neural network to understand symmetry and use it as a feature extractor to greatly improve the understanding of images.” , Says Agar, the lead author of this paper.Description: “Symmetry-aware recursive image similarity search for material microscopes”, released today npj calculation material.. In addition to agar, authors include Tri NM Nguyen from Lehigh University, Yichen Guo, Shuyu Qin, Kylie S. Frew, and Ruijuan Xu from Stanford University. The lead author, Nguyen, is an undergraduate student at Lehigh University and currently holds a PhD. At Stanford.
The team was able to reach the projection by adopting the nonlinear dimensionality reduction techniques Uniform Manifold Approximation and Projection (UMAP). This approach allows researchers to “… learn the topology and high-level structure of data in an ambiguous way and compress it into 2D.”
“When we train a neural network, the result is a vector, or a series of numbers that are compact descriptors of features. These features help us classify things so that similarities are learned.” Agar says. “However, there may be more than 512 different features, so what is generated is still quite large in space, so it needs to be compressed into a human-readable space such as 2D, 3D, etc. Maybe, 4D. “
This allowed Agar and his team to take over 25,000 images and group materials of very similar classes.
“Similar types of structures in materials are semantically close, and certain trends can be observed, especially when some metadata filters are applied,” says Agar. “When you start filtering by who made the deposit, who made the material, what they were trying to do, what the material system was, it actually started refining and gaining more and more similarity. You can link its similarities to other parameters such as properties. “
This work shows how improved data storage and management can accelerate material discovery quickly. According to agar, of particular value are the images and data produced by the failed experiments.
“No one publishes the results of the failure. A few years later, it’s a huge loss because someone repeats the same series of experiments,” says Agar. “So you waste really good resources on experiments that probably won’t work.”
Instead of losing all that information, the data already collected can be used to generate new trends never seen before, dramatically speeding up discoveries, says Agar. ..
This study is the first “use case” of an innovative new data storage company housed at Oak Ridge National Laboratory. DataFed.. DataFed, according to its website, “… Federation, Big-Data storageA full lifecycle management system for computational science and / or data analysis within a distributed high performance computing (HPC) and / or cloud computing environment. ”
“My team at Lehi was part of the design and development of DataFed in terms of making it relevant to scientific use cases,” says Agar. “Lehi is the first live implementation of this fully scalable system. It’s a federated database, so anyone can pop up their server and connect to a central facility.”
Agar is a machine learning expert on the Lehigh University Presidential Nanohuman Interface Initiative Team. An interdisciplinary initiative that integrates social science and engineering seeks to transform the way humans interact with the means of scientific discovery to accelerate innovation.
“One of the key goals of Lehi’s Nano / Human Interface Initiative is to put relevant information at the fingertips of the experimenter, enabling more informed decision making and accelerating scientific discoveries. It’s about providing information, “says Agar. “Human memory and memory capacity is limited. DataFed is a modern Memex that provides a memory of scientific information that is easy to find and recall.”
DataFed provides researchers engaged in interdisciplinary team science with a particularly powerful and valuable tool, giving researchers collaborating on different / remote team projects access to each other’s raw data. will do so. This is one of the key components of the Lehi Presidential Nano / Human Interface (NHI) initiative to accelerate scientific discoveries, “said Lehi’s Professor of Materials Science and Engineering, Alcoa Foundation and Director of Nano / Human Interface. Martin P. Harmer said. Initiative.
Symmetry-aware recursive image similarity search for material microscopes, npj calculation material, DOI: 10.1038 / s41524-021-00637-y
Quote: Research on speed materials, a new neural network for understanding symmetry (October 8, 2021), from https: //phys.org/news/2021-10-neural-network-symmetry-materials.html Acquired on October 8, 2021.
This document is subject to copyright. No part may be reproduced without written permission, except for fair transactions for personal investigation or research purposes. The content is provided for informational purposes only.
New neural networks for understanding symmetry, speeding up materials research
https://phys.org/news/2021-10-neural-network-symmetry-materials.html New neural networks for understanding symmetry, speeding up materials research