These are preliminary reports that have not been peer-reviewed. They should not be regarded as conclusive, guide clinical practice/health-related behavior, or be reported in news media as established information. For more information, please see our FAQs.
autoencoder_chemrxivsubmit1.pdf (3.14 MB)
Machine Learning of Optical Properties of Materials - Predicting Spectra from Images and Images from Spectra
Preprints are manuscripts made publicly available before they have been submitted for formal peer review and publication. They might contain new research findings or data. Preprints can be a draft or final version of an author's research but must not have been accepted for publication at the time of submission.
submitted on 10.07.2018 and posted on 11.07.2018by Helge S. Stein, Dan Guevarra, Paul F Newhouse, Edwin Soedarmadji, John Gregoire
As the materials science community seeks to capitalize on recent advancements in computer science, the sparsity of well-labelled experimental
data and limited throughput by which it can be generated have inhibited deployment of machine learning algorithms to date. Several successful
examples in computational chemistry have inspired further adoption of machine learning algorithms, and in the present work we present
autoencoding algorithms for measured optical properties of metal oxides, which can serve as an exemplar for the breadth and depth of data
required for modern algorithms to learn the underlying structure of experimental materials science data. Our set of 180,902 distinct materials
samples spans 78 distinct composition spaces, includes 45 elements, and contains more than 80,000 unique quinary oxide and 67,000 unique
quaternary oxide compositions, making it the largest and most diverse experimental materials set utilized in machine learning studies. The
extensive dataset enabled training and validation of 3 distinct models for mapping between sample images and absorption spectra, including a
conditional variational autoencoder that generates images of hypothetical materials with tailored absorption properties. The absorption patterns
auto-generated from sample images capture the salient features of ground truth spectra, and direct band gap energies extracted from these
auto-generated patterns are quite accurate with a mean absolute error of 240 meV, which is the approximate uncertainty from traditional
extraction of the band gap energy from measurements of the full transmission and reflection spectra. Optical properties of materials are not only
ubiquitous in materials applications but also emblematic of the confluence of underlying physical phenomena that yield the type of complex data
relationships that merit and benefit from neural network-type modelling.