Advanced Deep Learning Methods for Chemistry and Material Science

Abstract

In chemistry and material science, scientific discovery is usually achieved through a combination of wet-lab experiments and first-principle computational methods. These traditional approaches are often time-consuming and computationally expensive, significantly slowing down the pace of discovery. In recent years, researchers have started exploring deep learning methods to accelerate this process and reduce the cost. While these initial attempts have shown great promise, there remains significant challenges that must be addressed to fully realize the potential of deep learning in this field. This dissertation advances research on deep learning methods for molecular and material property predictions from three key perspectives: 1) Molecular Representation Learning: We propose an expressive neural network (HMGNN) that can learn better molecule representations and achieves state-of-the-art performance in molecular property prediction tasks. 2) Multi-modal Molecular Learning: we develop a retrieval augmentation method (RTMol) that leverages additional information present in scientific literature to augment molecular structures for accurate property prediction. 3) Label Efficiency: we propose two methods to effectively train neural networks for material and molecular property prediction with limited labeled data. The first method utilizes materials labeled by computationally efficient labeling methods to augment the limited labeled training data. The second method (DSP) selects task specific pre-training subsets to effectively adapt already pre-trained neural networks to downstream tasks.