quantummind / deepq

Quantum circuit optimization with deep learning

Quantum circuit optimizer

A. Zlokapa and A. Gheorghiu, “A deep learning model for noise prediction on near-term quantum devices.” https://arxiv.org/abs/2005.10811

A. Zlokapa and A. Gheorghiu, “A deep learning approach to noise prediction and circuit optimization for near-term quantum devices.” IEEE/ACM International Conference on High Performance Computing, Networking, Storage and Analysis, November 2019. 1st place, ACM SRC SC19. 2nd place, ACM SRC International.

Abstract

We present an approach for a deep-learning compiler of quantum circuits, designed to reduce the output noise of circuits run on a specific device. We train a convolutional neural network on experimental data from a quantum device to learn a hardware-specific noise model. A compiler then uses the trained network as a noise predictor and inserts sequences of gates in circuits so as to minimize expected noise. We tested this approach on the IBM 5-qubit devices and observed a reduction in output noise of 12.3% (95% CI [11.5%, 13.0%]) compared to the circuits obtained by the Qiskit compiler. Moreover, the trained noise model is hardware-specific: applying a noise model trained on one device to another device yields a noise reduction of only 5.2% (95% CI [4.9%, 5.6%]). These results suggest that device-specific compilers using machine learning may yield higher fidelity operations and provide insights for the design of noise models.

Overview

• main.ipynb: Summary of results, comparing compilation with the deep learning noise model to the IBM Q compiler.
• dd_analysis.py: Demonstrates that learned sequences do not achieve dynamical decoupling to first order, despite reducing noise more than dynamical decoupling.

Data

To run the code and notebooks, unzip the data available at https://caltech.box.com/s/jni52ra5qpq28f9tunaob9auwvisavb9 and add it directly to the main directory of the project.

Circuit generation

• generate_circuits.py: Helper functions for making supremacy-style circuits with a layer of random single-qubit gates from {sqrt(X), sqrt(Y), sqrt(W)} followed by a CX gate between any two qubits. Generation of training set, supremacy_all_5_unique/circuits.npy.
• make_test_circuits.py: Generation of test circuits, test_circuits_5.npy.

Deep learning

• model.py: Definition of neural network and circuit pair dataset structure.
• train.py: Train model on training set with 80% training set, 10% validation set (for early stopping), and 10% test set.
• test_compiler.py: Compile test circuits with maximum IBM Q compiler optimization and then pad the output of that with the deep learning model.
• result_plotter.ipynb: Evaluate model performance on training/validation/test set.

Circuit Evaluation

• run_circuits.py: Run training set circuits, saved in supremacy_all_5_unique/burlington_noise.npy for Burlington and similarly for London.
• run_test_compiled.py: Run test set circuits, including free evolution and compiled, saved in test_noise_5.

Citation

@misc{alex2020deep,
    title={A deep learning model for noise prediction on near-term quantum devices},
    author={Alexander Zlokapa and Alexandru Gheorghiu},
    year={2020},
    eprint={2005.10811},
    archivePrefix={arXiv},
    primaryClass={quant-ph}
}