About this Abstract |
Meeting |
1st World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022)
|
Symposium
|
First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022)
|
Presentation Title |
Efficient, Interpretable Atomistic Graph Neural Network Representation for Angle-Dependent Properties and its Application to Optical Spectroscopy Prediction |
Author(s) |
Tim Hsu, Nathan Keilbart, Stephen Weitzner, James Chapman, Tuan Anh Pham, Roger Qiu, Xiao Chen, Brandon Wood |
On-Site Speaker (Planned) |
Tim Hsu |
Abstract Scope |
Graph neural networks have gained popularity for learning properties of molecular and atomic structures thanks to their physically informed representation of atoms and bonds. However, standard GNNs lack explicit information about bond angles, which impact electronic structure and chemical hybridization. A recent formulation (ALIGNN) uses auxiliary line graphs to explicitly represent bond angles. In this work, we extend ALIGNN to also capture dihedral angles and chirality (ALIGNN-d). Such encoding is particularly important for amorphous/disordered systems and functional material design in phase spaces where angular information is critical. Using optical absorption spectroscopy of copper aqua complexes as a demonstration, we show that ALIGNN-d can accurately and efficiently predict key spectral features. Further, the expressive power of ALIGNN-d is shown to match the far less scalable maximally-connected graph, in which all pairwise bonds are explicitly encoded. Lastly, we show how the physical motivations underlying ALIGNN-d can be leveraged for model interpretability. |
Proceedings Inclusion? |
Undecided |