Research
Machine Learning of Materials’ Optical Properties
Stein, H., Guevarra, D., Newhouse, P., Soedarmadji, E., Gregoire, J. Machine learning of optical properties of materials – predicting spectra from images and images from spectra. Chemical Sciences, DOI: 10.1039/C8SC03077D (2018).
Scientific Achievement
JCAP researchers created a model that predicts the optical absorption spectrum from a material’s image and vice versa using the world’s largest materials image and spectroscopy dataset.
Significance & impact
Generative models learn data relationships that enable predictions in unexplored spaces, and this seminal demonstration in experimental materials science provides guidance on how to deploy machine learning for materials discovery.
Research Details
Images and UV-Vis absorption spectroscopy for 178,994 metal oxide samples enabled deep neural network training
The model correctly predicts features such sub-gap absorption from only a materials image, while a human cannot
The autoencoder latent space enables prediction of additional properties, such as band gap energy, from a material’s image.
Contact: stein@caltech.edu, gregoire@caltech.edu
Reprinted from Stein, H., Guevarra, D., Newhouse, P., Soedarmadji, E., Gregoire, J. Machine learning of optical properties of materials – predicting spectra from images and images from spectra. Chemical Sciences, DOI: 10.1039/C8SC03077D (2018).
Schematic visualization of the 3 types of learning models for optical properties of materials.
Reprinted from Stein, H., Guevarra, D., Newhouse, P., Soedarmadji, E., Gregoire, J. Machine learning of optical properties of materials – predicting spectra from images and images from spectra. Chemical Sciences, DOI: 10.1039/C8SC03077D (2018).
Varying the size of the model training set demonstrates that large datasets are critical for creating a predictive model.