Fig. 2: Forward property prediction using semi-supervised deep learning.
From: Attribute driven inverse materials design using deep learning Bayesian framework
Properties information embedded in molecular structures is extracted using unsupervised learning of DBN, as shown in a. The Pearson’s correlation coefficient between activation probability of each node and the properties is shown in the figure. The DBN network used in this paper is shown in the rightmost part of the figure. Highest correlation for each layer output is shown in the leftmost part of the figure. Similarly, correlation between the properties is shown in b. The quantile-quantile plot for deep learning predicted properties is shown in c. Violin plot in d shows comparison of the PDF for predicted properties and dataset. Kulback–Liebler (KL) divergence between the PDFs is also shown in the figure.
