A team of researchers from MIT, the University of Massachusetts at Amherst, and the University of California at Berkeley have devised a computer learning method of automatically scanning and compiling recipes for new materials based on certain requirements.
The team has released a paper in the latest issue of the Chemistry of Materials journal highlighting the process, and hope that it will bring a touch of well-needed automation to the field of materials science. Until now, any researchers involved with developing new materials would have to rely on a combination of expertise, intrinsic knowledge and a lot of reading if they wanted to develop a new material. Those days of laboriously flicking through research papers may soon be numbered.
“Computational materials scientists have made a lot of progress in the ‘what’ to make—what material to design based on desired properties,” said Elsa Olivetti, assistant professor of Energy Studies in MIT’s Department of Materials Science and Engineering. “But because of that success, the bottleneck has shifted to, ‘Okay, now how do I make it?’”
The goal of the team is to create a database that populates itself from existing electronic research literature, and make that database available to other researchers around the world. Then, the researchers would simply add the requirements based on precursor materials, or manufacturing processes, for example, and the system would display recommendations for the new material recipe.
The system can scan through research papers and group relevant words and ingredients based on their relevance, and can also infer data based on material properties such as the range of temperature required for the target material’s fabrication.
The early system has progressed with a combination of supervised and unsupervised learning. In the former method, the program requires a human supervisor to edit the test data in order to make it readable by machine. In the latter method, the data is not truncated and the machine is allowed to process the data and group it automatically (which more closely resembles a real-life scenario).
During early trials of the system, the research team was only able to trawl through about 100 papers due to the need to manually edit the test data. But since employing a computer-learning algorithm called “Word2vec,” the team has been able to increase that figure dramatically to around 640,000 papers.
In trials using the truncated data sets, the team was able to identify the paragraphs that contained recipes with an accuracy of 99 percent, and was able to label the words in those paragraphs with an accuracy of 89 percent.
“This is landmark work," said Ram Seshadri, professor of Materials Science at the University of California at Santa Barbara. “The authors have taken on the difficult and ambitious challenge of capturing, through AI methods, strategies employed for the preparation of new materials. The work demonstrates the power of machine learning, but it would be accurate to say that the eventual judge of success or failure would require convincing practitioners that the utility of such methods can enable them to abandon their more instinctual approaches."
Now the team hopes to improve the method’s accuracy by employing deep learning techniques to the system, which, when deployed, will hopefully make light work of materials development—or at least lighter.