Neural network accelerator for quantum control

David Xu, A. Baris Özgüler, Giuseppe Di Guglielmo, Nhan Tran, Gabriel N. Perdue, Luca P. Carloni, Farah FahimarXiv 2022 · DOI · OpenReview

TL;DR

This paper presents a novel approach to quantum control by using a neural network to predict optimal control pulses, which is then synthesized into a hardware accelerator on an FPGA. This method offers a more efficient and cost-effective solution for quantum control, especially for deployment in resource-constrained or cryogenic environments.

What Problem Does This Paper Address?

Quantum computers, while promising, are currently limited by errors that hinder their practical use. Quantum control aims to precisely manipulate qubits to improve performance. However, traditional methods for determining optimal control parameters are computationally intensive, making them impractical for real-time use, especially in the harsh environments of quantum computing hardware (like cryogenic temperatures). Existing hardware solutions, such as lookup tables, are often imprecise and costly. This paper addresses the need for an efficient, low-latency, and cost-effective quantum control solution that can operate close to the quantum hardware.

Methods

The core of this work involves training a machine learning model, specifically a multi-layer perceptron (MLP), to approximate the results of traditional quantum optimal control algorithms. The trained model is then synthesized into a hardware accelerator using the hls4ml framework, which is designed for deploying ML models on FPGAs and ASICs with a focus on low power and low latency. The authors explored different MLP architectures and quantization strategies to balance accuracy, resource utilization, and performance. They also investigated the trade-offs between different FPGA development boards with varying resource constraints.

Datasets

The primary dataset used for training and evaluation was generated using Juqbox.jl, an open-source software package for solving quantum optimal control problems. This dataset consists of samples where each sample includes a rotation angle ($\beta$) and the corresponding 20-dimensional vector of pulse parameters ($\alpha$) required to achieve that rotation. The authors specifically removed samples associated with $\beta = -\pi$ to ensure consistency in the dataset.

Key Insights and Ideas

  • ML for Control Parameter Prediction: The paper demonstrates that machine learning can effectively learn the complex relationship between desired quantum gate operations and the precise control pulses needed to achieve them, offering a significant speedup over traditional methods.
  • Hardware Acceleration for Quantum Control: Synthesizing the ML model into an FPGA-based accelerator allows for low-latency, high-throughput quantum control directly at the hardware level, reducing the need for external computation and data transfer.
  • Resource Optimization for Cryogenic Environments: The focus on lightweight ML models and efficient FPGA implementation makes this approach suitable for deployment in resource-constrained and cryogenic environments where traditional computing hardware is not feasible.
  • Trade-offs in FPGA Implementation: The study highlights the importance of balancing model complexity, quantization, and hardware resource utilization to achieve desired performance metrics on FPGAs.
  • "Smoothed" ML Output: The ML model's output, when compared to the optimizer's output, appears "smoothed," suggesting it can effectively filter out random noise from the optimization process while still capturing the essential control dynamics.

Key Results

  • The developed ML model achieves a gate fidelity of over 0.99 for predicting optimal pulse parameters.
  • The synthesized FPGA implementation exhibits a latency of 175 ns and a pipeline initiation interval of 1 cycle.
  • The implementation fits on a low-resource FPGA (Digilent Genesys 2), utilizing 54% of available DSPs and 13% of LUTs for the initial implementation.
  • Further model shrinking and aggressive quantization allowed the implementation to fit on an even smaller board (Avnet Ultra96-V2), using 41% of DSPs and 58% of LUTs.
  • The final implementation on an Artix-7 chip achieved a fidelity above 0.99 for all tested $\beta$ values.

Limitations and Future Work

  • Fidelity Gap: While the ML model achieves >0.99 fidelity, the training data itself reaches >0.999 fidelity. The authors suggest exploring techniques like AutoML (e.g., AutoKeras) and investigating larger models to bridge this gap.
  • Branched Models: The paper hints at the potential of branched neural network architectures, where each output parameter is handled by a dedicated branch, to potentially improve performance.
  • Lookup Table Comparison: The authors acknowledge that lookup tables might be simpler for very specific gates but argue that neural networks offer better scalability and interpolation capabilities for more complex scenarios. Further comparison of accuracy with linear interpolation methods is suggested.
  • Multi-Qubit Systems: Extending the approach to control multi-qubit systems and more complex quantum operations is a significant future direction.
  • Extreme Low-Temperature Performance: Evaluating the performance of the accelerator in extreme low-temperature environments is a crucial next step.
  • Custom eFPGA Systems: Designing custom eFPGA-based systems for even greater optimization is also proposed.

Summary updated

Abstract

Efficient quantum control is necessary for practical quantum computing implementations with current technologies. Conventional algorithms for determining optimal control parameters are computationally expensive, largely excluding them from use outside of the simulation. Existing hardware solutions structured as lookup tables are imprecise and costly. By designing a machine learning model to approximate the results of traditional tools, a more efficient method can be produced. Such a model can then be synthesized into a hardware accelerator for use in quantum systems. In this study, we demonstrate a machine learning algorithm for predicting optimal pulse parameters. This algorithm is lightweight enough to fit on a low-resource FPGA and perform inference with a latency of 175 ns and pipeline interval of 5 ns with $~>~$0.99 gate fidelity. In the long term, such an accelerator could be used near quantum computing hardware where traditional computers cannot operate, enabling quantum control at a reasonable cost at low latencies without incurring large data bandwidths outside of the cryogenic environment.
BibTeX
@misc{xu2022neural,
  author = {David Xu and A. Baris Özgüler and Giuseppe Di Guglielmo and Nhan Tran and Gabriel N. Perdue and Luca P. Carloni and Farah Fahim},
  title = {{Neural network accelerator for quantum control}},
  year = {2022},
  eprint = {2208.02645},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.02645},
}
Figure