Following their proposal of a neural network-based artificial intelligence (AI) system that sifts through scientific papers and finds "recipes" for specific types of materials, researchers at the Massachusetts Institute of Technology say they have developed an AI that identifies consistent higher-level patterns across recipes.
The algorithm relies on statistical methods providing a natural mechanism for generating original recipes, and the team uses it to suggest alternative recipes for known materials. The variational autoencoder network is designed to refine input vectors into smaller vectors whose numbers are significant for every input, so it has a middle layer with only a few nodes. Once the network was trained on recipes for manganese dioxide and related compounds, the team built a map depicting the values the middle nodes assumed for each example in the training set.
The researchers note the AI recognized correlations between precursor chemicals used in materials recipes and the crystalline structures of the resulting products.
From MIT News
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found