Deep Neural Network Assisted Quantum Chemistry Calculations on Quantum Computers

30 October 2023, Version 1
This content is a preprint and has not undergone peer review at the time of posting.


The variational quantum eigensolver (VQE) is a widely employed method to solve electronic structure problems on the current noisy intermediate-scale quantum (NISQ) devices. However, due to inherent noise in the NISQ devices, VQE results on NISQ devices often deviate significantly from the results obtained on noiseless statevector simulators or traditional classical computers. The iterative nature of the VQE further amplifies the errors in each loop. Recent works have explored ways to integrate deep neural networks (DNN) with VQE to mitigate the iterative errors, albeit, primarily limited to the noiseless statevector simulators. In this work, we trained DNN models across various quantum circuits and examined the potential of two DNN-VQE approaches, DNN1 and DNNF, for predicting the ground state energies of small molecules in the presence of device noise. We carefully examined the accuracy of the DNN1, DNNF, and VQE methods on both noisy simulators and real quantum devices by considering different ansatzes of varying qubit counts and circuit depths. Our results illustrate the advantages and limitations of both VQE and DNN-VQE approaches. Notably, both DNN1 and DNNF methods consistently outperform the standard VQE method in providing more accurate ground-state energies in noisy environments. However, despite being more accurate than VQE, the energies predicted using these methods on real quantum hardware remain meaningful only at reasonable circuit depths (depth = 15, gates = 21). At higher depths (depth = 83, gates = 112), they deviate significantly from the exact results. Additionally, we find that DNNF does not offer any notable advantage over VQE in terms of speed. Consequently, our study recommends DNN1 as the preferred method for obtaining quick and accurate ground state energies of molecules on the current quantum hardware, particularly for quantum circuits with lower depth and fewer qubits.


Quantum computing
Computational chemistry
Machine learning
Neural networks
Error mitigation

Supplementary materials

Supporting Information
Effect of parameter initialization for the LiH molecule, details of the employed quantum circuits, DNN model layout, time taken to compute the energies on a noisy simulator, accuracy of various methods on a statevector simulator, comparison of DNNF and MP2-init-VQE results on a noisy simulator, convergence on a statevector simulator, variation of energies with number of iterations a noisy simulator, and deviation of various methods from the CCSD energies.


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.