Pure Quantum Generators for Molecular Generation with Quantum GANs : A quantum-classical comparative study

Sammanfattning

Navigating chemical space for de novo drug design is a computationally expensive task. Data-driven methods, like those based on machine learning, have shown great potential for accelerating the drug discovery process. By utilizing a high degree of flexibility and leveraging large amounts of chemical data, effective machine-learning approaches have been achieved. Quantum computing, especially in the context of machine learning, represents a novel avenue for data-driven methods. Using principles of quantum mechanics, such as superposition and entanglement, so-called quantum neural networks (QNNs) can demonstratively outperform their classical counterparts in terms of expressive power. While quantum machine learning has limited practical applications due to the state of modern quantum devices, exploring such quantum algorithms is still important in preparation for more capable quantum hardware. This thesis compares the performance of a quantum generative adversarial network (qGAN) with a classical GAN baseline. Experimental results revealed that while the qGAN architecture demonstrated potential, it did not surpass the classical baseline across most metrics. However, the quantum models utilized significantly fewer parameters, indicating potential for parameter efficiency. The findings highlight the importance of overcoming optimization challenges to fully realize the advantages of qGANs in molecule generation tasks.