Physics Simulation Via Quantum Graph Neural Network

TL;DR

This paper introduces two quantum graph neural network models for particle interaction simulation, comparing their performance with classical models and highlighting potential advantages of quantum approaches.

Contribution

The paper presents two realizations of quantum graph neural networks for particle simulation, demonstrating their potential advantages over classical models and exploring hyperparameter tuning effects.

Findings

Implementable QGNN shows potential advantage over CGNN.

All models rapidly reduce loss but are moderately inaccurate.

Hyperparameter tuning improves classical GNN accuracy.

Abstract

We develop and implement two realizations of quantum graph neural networks (QGNN), applied to the task of particle interaction simulation. The first QGNN is a speculative quantum-classical hybrid learning model that relies on the ability to directly utilize superposition states as classical information to propagate information between particles. The second is an implementable quantum-classical hybrid learning model that propagates particle information directly through the parameters of $RX$ rotation gates. A classical graph neural network (CGNN) is also trained in the same task. Both the Speculative QGNN and CGNN act as controls against the Implementable QGNN. Comparison between classical and quantum models is based on the loss value and accuracy of each model. Overall, each model had a high learning efficiency, in which the loss value rapidly approached zero during training; however, each model was moderately inaccurate. Comparing performances, our results show that the Implementable QGNN has a potential advantage over the CGNN. Additionally, we show that a slight alteration in hyperparameters in the CGNN notably improves accuracy, suggesting that further fine tuning could mitigate the issue of moderate inaccuracy in each model.