Machine Learning in the Quantum Regime Through Physical-Principle-Informed Representations

Materials design in the quantum regime call for the integration of multi-tier materials information that goes beyond atomic structures. Especially, many quantum behaviors are greatly controlled by local bonding environments and physical constraints related to symmetry. In this talk, I will give several examples of how domain knowledge and physical principles for quantum material systems can be incorporated into machine learning frameworks through representation learning to greatly improve the performance of machine learning models for property predictions. Motivated by Pauling’s rules, I will show that local bonding environments (structure motifs) can be incorporated into a graph-based machine-learning architecture to make reliable property predictions for solid-state quantum materials including complex metal oxides. The proposed atom-motif dual network model demonstrates the feasibility to incorporate beyond-atom materials information in a graph network framework and achieves state-of-the-art performance in predicting the electronic structure properties of complex metal oxides. Through unsupervised learning, abstracted material information such as chemical formulas and motif connections can be combined with national language processing technologies to effectively represent fundamental elements in materials and use them in downstream learning tasks. I will demonstrate how contrastive representation learning can be used to incorporate physical constraints that control the collective behavior of electron densities into neural-network-based density functional design. At the end of the talk, I will discuss the continued development of machine learning models for quantum materials that embrace local/global symmetries and symmetry-based interactions.