Rotation- and Permutation-Equivariant Quantum Graph Neural Network for 3D Graph Data.

Area of Science:

• Artificial Intelligence
• Quantum Computing
• Computational Chemistry

Background:

• Existing equivariant quantum graph neural networks (EQGNNs) primarily consider permutation symmetry, limiting their effectiveness with 3D graph data.
• Failure to fully exploit geometric and non-geometric information leads to suboptimal performance in current EQGNN models for complex 3D graph processing.

Purpose of the Study:

• To develop a novel quantum graph neural network that incorporates both rotation and permutation equivariance for enhanced 3D graph data processing.
• To address the limitations of existing EQGNNs by extracting both geometric and non-geometric features more effectively.

Main Methods:

• Derivation of constraints for rotation and permutation equivariance.
• Proposal of a novel rotation- and permutation-equivariant quantum graph neural network (RP-EQGNN).
• Design of an equivariant module for geometric information extraction and a convolution-entanglement module for non-geometric information extraction.
• Implementation of an edge entanglement strategy to differentiate entanglement operations based on edge heterogeneity.

Main Results:

• RP-EQGNN demonstrates superior performance in graph regression on the QM9 and OC20 datasets, achieving lower Mean Absolute Error (MAE) compared to Q3DGL and EQC.
• The model achieves results comparable to state-of-the-art methods like EquiformerV2, Geoformer, SO3KRATES, and HEGNN.
• RP-EQGNN shows an advantage in point cloud classification on the ModelNet40 dataset over quantum models such as sQCNN-3D and PI-QSVM.

Conclusions:

• RP-EQGNN offers an innovative approach for processing 3D graph data by effectively leveraging rotation and permutation symmetries.
• The developed model establishes a foundation for future research into symmetries within graph neural networks, particularly for complex 3D applications.