Practical application of quantum neural network to materials informatics.

Area of Science:

• Quantum Computing
• Materials Science
• Machine Learning

Background:

• Quantum neural networks (QNNs) offer high expressibility and resistance to overfitting, making them suitable for materials informatics (MI) challenges with limited data.
• The application of QNNs to multivariate regression tasks, particularly in MI, remains underexplored, with limited understanding of model construction.

Purpose of the Study:

• To construct and evaluate a QNN model for predicting the melting points of metal oxides, serving as a multivariate regression task in MI.
• To investigate the impact of different QNN architectures, including encoding methods and entangler arrangements, on model performance.

Main Methods:

• Exploration of various QNN architectures, focusing on encoding strategies and entangler configurations.
• Assessment of circuit depth, width, and entangler type ('linear') for optimal QNN performance.
• Comparison of QNN model generalization against classical neural network (NN) models.

Main Results:

• Shallow-depth QNNs with sufficient entanglement demonstrated adequate expressibility.
• A 'linear' entangler arrangement proved effective for achieving necessary entanglement.
• Increasing circuit width enhanced QNN model expressibility and generalization performance.
• QNN models outperformed classical NN models in generalization and showed no overfitting with a well-designed encoder.

Conclusions:

• QNNs are a viable and powerful tool for materials informatics, particularly for multivariate regression tasks like predicting material properties.
• Optimized QNN architectures, balancing expressibility and entanglement, can achieve superior performance compared to classical models.
• The findings pave the way for broader adoption of QNNs in accelerating materials discovery and design.