Succeeded in finding metallic glass 200 times faster by "machine learning"
Machine learning techniques are widely used in various fields such as artificial intelligence (AI) development and image recognition technology. Meanwhile, attempts have been made to explore unprecedented discoveries by using machine learning in the scientific world, and American researchersMetallic glassWe have succeeded in finding the synthetic pattern of 100 times faster than the conventional pattern.
Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments | Science Advances
Artificial intelligence accelerates discovery of metallic glass - Northwestern Now
There is no regularity in the element sequenceAmorphous metalAmong them, metallic glass causing glass transition is thought to have high corrosion resistance and wear resistance, it is thought that it can realize a material with high rigidity and lighter than steel, and is expected to be a next generation material. However, since metallic glass was discovered to be produced, the combination of metallic production found in 50 years is limited, and only a few kinds are expected to be put to practical use. This is because there are millions of combinations of metals and it is enormous when including up to that formulation, so even if we narrow down to promising candidates, we need to do a lot of experiments.
The research group of Northwestern University, SLAC National Accelerator Research Institute, NIST succeeded in quickly creating and screening hundreds of sample materials, incorporating machine learning for the production of metallic glasses, and three new types We have discovered that the metallic glass has been discovered. The research paper is published in Science Advances.
Dr. Amba Meta, co-author of this thesis, said, "The unique point of our approach is to use measurement results quickly It is a point where we make predictions and reflect the repeated results in the next machine learning process and experiments. "
In the research, he took in the data of 6000 metal glass production experiments collected over the past 50 years using the machine learning algorithm created by Mr. Logan Ward of the graduate student of Wolverton Institute. Based on what the machine learning algorithm learned firstly, sample alloys are generated in two different ways, the results of X-ray scanning of this alloy are used as a database, new machine learning is performed, and another sample is synthesized The process took place. Every time through the process, it seems that the proportion of discovering one metallic glass from 300 to 400 materials seems to be able to discover one metallic glass from two to three materials. By incorporating machine learning, metal We report that the speed required to discover glass is increased by about 200 times compared with the conventional method.
In particular, it is important that machine learning algorithms adopted in the experiment do not need to understand the conventional "theory", and the machine learning system created by the research group can be applied to other research. Human researchers can concentrate on other tasks that require their own intuition and creativity by releasing humans from non-creative experimental processes which human beings had no choice but to date, science It seems that Dr. Meta thinks that it may give a good influence to the entire community.