University of Ottawa

Purpose

The field of novel materials discovery has witnessed several interesting techniques that span across traditional synthetic chemistry and computational methods. Synthetic chemistry methods are normally Edisonian-based (i.e., trial-and-error) and often will rely on experiential knowledge to synthesize chemically stable materials. Machine Learning (ML) approaches, that are based on Deep Generative Modeling (DGM), are commendable alternatives for novel materials discovery due to their impressive ability in analyzing robust chemical design space. Famous for their speed, reliability, and low cost, DGM techniques can intelligently identify hidden patterns and correlations in a training dataset by solving an inverse design scheme. However, deep generative ML (DGML) approaches face challenges related to lattice reconstruction at the decoding phase, potentially leading to two major shortcomings. To address the highlighted challenges, the project will develop progressive deep learning approaches for novel materials discovery in two stages. The first stage will leverage on the technical strengths of both a semi-supervisory VAE (i.e. SS-VAE) model and an auxiliary GAN (i.e. A-GAN) model. The model architecture is referred to as Lattice-Constrained Materials Generative