Machine Learning Assisted Construction of a Shallow Depth Dynamic Ansatz for Noisy Quantum Hardware

Sonaldeep Halder,Anish Dey,Chinmay Shrikhande,Rahul Maitra

DOI: https://doi.org/10.1039/d3sc05807g

IF: 8.4

2024-01-18

Chemical Science

Abstract:The development of various dynamic ansatz-constructing techniques has ushered in a new era, rendering the practical exploitation of Noisy Intermediate-Scale Quantum (NISQ) hardware for molecular simulations increasingly viable. However, such ansatz construction protocols exhibit substantial measurement costs during their execution. This work involves the development of a novel protocol that capitalizes on regenerative machine learning methodologies and many-body perturbation theoretic measures to construct a highly expressive and shallow ansatz within the variational quantum eigensolver (VQE) framework with limited measurement costs. The regenerative machine learning model used in our work is trained with the basis vectors of a low-rank expansion of the N-electron Hilbert space to identify the dominant high-rank excited determinants without requiring a large number of quantum measurements. These selected excited determinants are iteratively incorporated within the ansatz through their low-rank decomposition. The reduction in the number of quantum measurements and ansatz depth manifest in the robustness of our method towards hardware noise, as demonstrated through numerical applications. Furthermore, the proposed method is highly compatible with state-of-the-art neural error mitigation techniques. This resource-efficient approach is quintessential for determining spectroscopic and other molecular properties, thereby facilitating the study of emerging chemical phenomena in the near-term quantum computing framework.

chemistry, multidisciplinary

What problem does this paper attempt to address?

### Problems the paper attempts to solve

This paper aims to solve the problems of high measurement cost and hardware - noise sensitivity when constructing a shallow - parameterized wave - function approximation (ansatz) in a dynamic variational quantum eigensolver (VQE) framework on current noisy intermediate - scale quantum (NISQ) hardware. Specifically, the authors propose a new method that combines regenerative machine - learning (ML) techniques and many - body perturbation theory (MBPT) to reduce the number of quantum measurements and the ansatz depth, thereby improving the method's robustness to hardware noise and enabling efficient calculation of molecular properties.

### Main contributions

1. **Reducing measurement cost**: By using a regenerative machine - learning model (such as the restricted Boltzmann machine RBM) to train the basis vectors of the low - rank - expanded N - electron Hilbert space, the dominant high - rank excited states are identified without a large number of quantum measurements.
2. **Reducing ansatz depth**: The selected excited states are iteratively incorporated into the ansatz through their low - rank decomposition, reducing the ansatz depth.
3. **Enhancing robustness**: Reducing the number of quantum measurements and the ansatz depth makes the method more robust to hardware noise.
4. **Compatibility with neural error - mitigation techniques**: This method is highly compatible with the state - of - the - art neural error - mitigation techniques, further improving its effectiveness on NISQ devices.
5. **Resource - efficient**: This method can efficiently determine spectroscopic and other molecular properties in the short - term quantum - computing framework, promoting the study of emerging chemical phenomena.

### Method overview

1. **Initial wave - function construction**: Use a shallow decoupled unitary coupled - cluster singles and doubles (dUCCSD) ansatz to construct a low - level wave - function on a quantum device, and obtain the state \( |\Psi_{SD}\rangle \) by variational optimization of parameters.
2. **New ansatz generation**: Derive a new ansatz \( \hat{U}_{\text{dUCC}}^{\text{RBM}} \) from the probability distribution of \( |\Psi_{SD}\rangle \), and select excitation operators according to the correlation probability ranking.
3. **RBM training**: Use the computational basis - state bit - strings and their probabilities to train the RBM and learn the correlations in the wave - function.
4. **High - rank excitation generation**: The RBM generates new binary vectors, including high - rank excited states (such as triple, quadruple excited states), forming a secondary excitation subspace.
5. **Scattering - operator introduction**: Indirectly introduce high - rank excitations through scattering operators (scatterers) to reduce the ansatz depth.
6. **Iterative optimization**: Repeat steps 4 to 6 until the RBM no longer generates new high - rank configurations, and finally obtain the RBM - dUCC ansatz.

### Conclusion

This method provides a new way to construct an efficient and expressive ansatz on NISQ devices, significantly reducing the measurement cost and ansatz depth, improving the robustness to hardware noise, and providing a new tool for the calculation of molecular properties.