Deep Learning for Predicting Molecular Electronic Properties

Applications of novel materials have a significant positive impact on our lives. To search for such novel materials, material scientists traverse massive datasets of prospective materials identifying ones with favourable properties. Prospective materials are screened by studying a suitable spectra of these materials. Contemporary methods like high-throughput screening are very time consuming for moderately sized datasets.

Recently, deep learning algorithms have proven to be successful in modelling very complex functions like the mapping from image to text, use for image captioning and the mapping from text in one language to another, used for machine translation.

In this thesis, we propose deep learning methods which are able to predict molecular orbital energies and spectra, from only the charges and coordinates of constituent atoms of test molecules. Our proposed machine learning (ML) model surpassed the state-of-the-art in prediction accuracy of the molecular orbital energies and based on our literature review it is the first ML model to predict molecular spectra.