Hybrid quantum-classical convolutional neural network for astrophysical object classification.

Abstract

Classifying astrophysical bodies like stars, galaxies, and black holes plays an underlying role in understanding the evolution of the universe. Effective analysis of a huge number of astronomical objects captured by telescopes is a challenge. For the classification of astronomical objects, quantum computing coupled with artificial intelligence proposes a good solution. Quantum machine learning (QML) utilizes the concepts of quantum computing to process huge datasets with greater accuracy and efficiency. In the current study, quantum features extraction with a convolutional neural network (CNN), named the AstroNet model, is proposed for the classification of astronomical objects. Quantum features extraction involves the conversion of raw images of astrophysical objects into quantum states for further processing. This is achieved by encoding the values of pixels in the amplitudes of quantum states using qubits. The AstroNet model consists of two phases. We first construct a quantum circuit. Entanglement is introduced by randomlayers, which employ cnot gates along with parameterized rotations. All quantum operations are implemented using the pennylane library in Python on a simulator. This enables efficient representation and processing of complex image data, setting the stage for quantum-powered classification in the second phase. These quantum features are passed to the proposed CNN model, which is designed from scratch, having seven layers, including two convolutional, two dense, two max pooling, and one flatten. The AstroNet is trained, based on the selected hyperparameters like the Adam optimizer, Sparse Categorical Cross-entropy, 32 batch-size, 0.0001 learning rate, and 10 epochs. The capability of the AstroNet model is examined based on the five top astrophysical object classification datasets including Galaxy Zoo, Galaxy Zoo 2, SpaceNet, Star-galaxy, and six classes of astronomical objects. The AstroNet model achieves up to 0.99 classification performance, which is much better than the previously reported methods in this field.