Quantum Generative Adversarial Autoencoders Achieve 0.02-0.06 Ha Accuracy Generating Quantum States with 6 Qubits

The creation of realistic and complex quantum states remains a significant challenge in the development of quantum technologies, and researchers are now exploring machine learning techniques to address this problem. Naipunnya Raj, Rajiv Sangle, and Avinash Singh, all from Fujitsu Research of India, along with Krishna Kumar Sabapathy, present a new model, the Quantum Generative Adversarial Autoencoder, which learns to generate quantum data efficiently. This innovative approach combines the strengths of autoencoders, which compress data into a manageable form, and generative adversarial networks, which learn the underlying patterns within that data, effectively giving the autoencoder the ability to create new, realistic quantum states. The team demonstrates the model’s capabilities by successfully generating both pure entangled states and accurately estimating the ground state energies of hydrogen and lithium hydride molecules, achieving errors of only 0. 02 and 0. 06 Hartree respectively, and paving the way for advancements in near-term quantum machine learning applications.

Quantum State Compression for Molecular Systems

This research details significant advances in quantum data compression and generative modeling, specifically for representing and creating quantum states of molecules. The core aim is to reduce the quantum resources needed for complex simulations. The team demonstrates that quantum data compression is achievable using QAEs, reducing the number of qubits required to represent quantum states.

Crucially, the research highlights that the rank of the density matrix, which describes the quantum state, determines the minimum number of qubits needed in the compressed representation. Combining autoencoding with generative adversarial training, through the development of a QGAA, proves particularly effective, leveraging the strengths of both techniques to improve the generation of realistic quantum states. The research demonstrates that directly training a QGAN on complex quantum states is less effective than first compressing the states with a QAE, suggesting that compression serves as a valuable preprocessing step for generative modeling. Experiments using hydrogen and lithium hydride molecules validate the performance of the models, employing metrics such as energy profile accuracy, fidelity, and reconstruction error to assess performance. This work addresses a key limitation of standard quantum autoencoders, which traditionally focus on data compression, by enabling them to generate new quantum data. The QGAA effectively learns and generates quantum data by first compressing quantum states into a lower-dimensional latent space and then utilizing an adversarial training framework to access the underlying quantum representation. Experiments involved training the QGAA to learn the latent representation of entangled two-qubit states and modeling the ground state energy profiles of hydrogen and lithium hydride molecules. Simulations achieved average errors of 0. 02 Hartree (Ha) for hydrogen and 0. 06 Ha for lithium hydride, demonstrating the potential of the method for accurate quantum chemistry calculations. The model combines an autoencoder, which compresses quantum states, with a generative adversarial network to learn the compressed state’s underlying structure, effectively giving the autoencoder the ability to generate new states. The researchers successfully generated both pure entangled states and the ground states of hydrogen and lithium hydride molecules. Simulations show average energy estimation errors of 0. 02 Hartree for hydrogen and 0.

06 Hartree for lithium hydride, up to six qubits. These results highlight the potential of this approach for quantum state generation and near-term quantum machine learning applications. The model offers a bidirectional advantage, enhancing the autoencoder’s generative capabilities while simultaneously reducing the computational resources needed for the generative adversarial network. Although the training process presented some challenges, the team demonstrated successful learning of the latent space and generation of quantum states closely resembling the target states. The generated states show promise as useful starting points for other quantum algorithms, potentially leading to faster convergence to optimal solutions, offering a promising path towards more efficient quantum simulations and calculations.