3 min Research Talk: Using Machine Learning for Materials Discovery and Property Prediction
Category
Published on
Abstract
Machine learning displays excellent potential for generating material property predictions and discovering novel compounds with desirable properties; however, it can be prohibitively costly to obtain data to train machine learning models. This barrier can be overcome by training models to predict related properties. To implement this strategy, we have created a computational tool that allows users to investigate how training models on multiple properties could lead to better prediction, even on datasets with missing property data. The framework for our model is an autoencoder, which consists of two parts: an encoder that projects molecules onto a continuous latent space and a decoder which decodes these points back into molecular structures. Our tool outputs principal component analysis plots to help users visualize how compounds are arranged within the latent space with respect to certain properties. With these functionalities, users can investigate how training on different sets of properties affects the organization of the continuous latent space. In turn, by customizing the organization of the latent space with respect to selected properties, users can discover novel compounds of potential application.
Bio
Mackinzie Farnell is an undergraduate student at Purdue University majoring in Materials Science and Engineering and Computer Science. She is interested in working at the intersection of these fields, and hopes to obtain a graduate degree in computational modelling of materials. In the summer of 2019, Machinzie did research with Professor Savoie from the Purdue Chemical Engineering School on using machine learning for materials discovery and property prediction.
Sponsored by
Cite this work
Researchers should cite this work as follows:
Time
Location
Rawls 1062, Purdue University, West Lafayette, IN