Discovery of new materials by AI makes the explosion speed & efficiency, both "A future where experiments with AI and robot alone will come"
ByUCL Mathematical & Physical Sciences
In the field of material science, for the past hundreds of years, scientists have discovered new materials due to factors such as trial and error, luck and coincidence. However, with the advent of artificial intelligence (AI), the process of discovering new materials is being skipped, so it is expected to make great progress in the future.
How AI is helping us discover materials faster than ever - The Verge
Past researchers in materials science used a method of mixing materials based on experiences and intuitions obtained in the past experiments and observing what is formed when creating new compounds It was. However, as of 2018, instead of using empirical knowledge and intuition, arithmetic processing by databases and computers has been used. It seems that we can predict which material can be produced by combining which materials, and it has become possible to efficiently produce the material.
The first example that utilized AI in the field of materials science is the study of Nicolas Marsali of the Swiss Federal Institute of Technology Lausanne et al. Mr. Marzaryi et al.GrapheneWe used a database of compounds to find materials like. Mr. Marzari and others announcedpaperAccording to the AI used in the research, we succeeded in narrowing down to about 2000 materials from more than 100,000 materials registered in the database.
Also, in a study published by a team such as Christopher Wolverton of Northwestern University,Discover how to accelerate the process up to 200 times faster using material AI to generate metallic glassDid. "AI's learning approach is similar to" how people learn new languages ", co-researcher Stanford University explains Apulva Meta. Usually, there are two ways to learn a new language: "Learning grammar rules" and "Listening to someone's story", but the AI used by the research team learned with these two approaches There is that.
In order to let AI learn, the research team first investigated recipes making various kinds of metallic glasses from past published papers, and incorporated the collected recipes as "grammar rules" into machine learning algorithms . Then, it said that this algorithm began to learn independently by which combination of materials produce new metallic glasses. According to Mr. Meter, "This algorithm's self-learning is similar to the way" I want to learn French, go to France and improve the language skills locally. "
Using AI makes it possible to synthesize and test thousands of materials at once, but it is not possible to test all the combinations of the periodic table, no matter how much time is needed Mr. Walberton is talking. In this research, the role of AI was "not to discover new materials, but to narrow down the scope that researchers investigate." For this reason, it does not support that machine learning is all-purpose, but by the proposal from AI, there is no doubt that the research team was able to discover new metallic glasses.
With the advent of AI, the future of materials science seems promising, but the challenges still remain. Mr. Marzali says, "There are errors in predicting computers, we will make predictions with a simple model that does not consider the real world." Originally, although the properties of compounds may change depending on temperature and humidity, it seems that there is a technical problem that these information can not be included in the machine learning algorithm being used.
In addition, Mr. Walberton points out that "Researchers do not have enough data on all compounds" as another issue. In other words, even if the machine learning algorithm is excellent, accurate results can not be obtained because of lack of data in the first place. However, if these problems can be completely overcome, "A human being will not participate in the experiment and the future will experiment with only AI and robots," Walberton said.