Synthesis of machine learning and physical principles for solid-state quantum materials  

Fundamental physical principles control the collective behavior of quantum matter. Solid-state material systems are encoded by multiple tiers of material information and their communications from a material informatics point of view are defined by local and global symmetries. The interactions within and between each tier of material information that resemble a quantum matter can be effectively represented by a machine learning architecture with physical principles incorporated. Funded by the DOE Early Career Program and the DOE EFRC center, the overarching objective of my research in this field is the creation of an enhanced learning framework for solid-state quantum matter by incorporating electron orbitals, site symmetries, and physical constraints in a synergistic way. The incorporation of physical principles in these learning tools hosts the potential to understand complex phenomena in large-scale material systems, discover new material knowledge, and accelerate the discovery and design of functional materials at an unprecedented rate.

Motif centric learning framework

As inspired by Pauling’s rule, we demonstrated that structure motifs in inorganic crystals can serve as a central input to a machine learning framework. We created a motif-centric learning framework by combining motif information with the atom-based graph neural networks. The atom-motif dual network (AMDNet) outperforms the state-of-the-art models in predicting the electronic structures of metal oxides. This work illustrates the route toward the fundamental design of machine learning architecture for complex materials by incorporating beyond-atom physical principles. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

H. R. Banjade, S. Hauri, S. Zhang, F. Ricci, W. Gong, G. Hautier, S. Vucetic, Q. Yan, “Structure motif centric learning framework for inorganic crystalline systems”, Sci. Adv., 7, eabf1754 (2021)

W. Gong, Q. Yan, “Graph-based deep learning frameworks for molecules and solid-state materials”, Comput. Mater. Sci. 195, 110332 (2021)

Figure4_updated.png

Atom2vec: learning from atoms

Motivated by the recent achievements of artificial intelligence (AI) in linguistics, in collaboration with Shou-Cheng Zhang’s group at Stanford, we designed a machine learning tool (Atom2Vec) to learn properties of atoms from materials data on its own. Atom2Vec realizes knowledge representation of atoms via unsupervised learning and could serve as a foundational step toward materials discovery and design fully based on machine learning.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Q. Zhou, P. Tang, S. Liu, J. Pan, Q. Yan, S. -C. Zhang, “Learning atoms for materials discovery”, PNAS 115, E6411 (2018).

PNAS figure 2.png