Machine Learning for Arbitrary Single-Qubit Rotations on an Embedded Device

TL;DR

The paperPresent a multi-stage ML workflow for arbitrary single-qubit rotations on embedded hardware near superconducting qubits, combining statevector-based bootstrapping, ARB-guided hardware fine-tuning, and a hardware-conscious deployment pipeline. It introduces Adapted Randomized Benchmarking to estimate non-Clifford gate fidelity from measurements, and demonstrates a compact neural network translating rotation angles into pulse coefficients, which is then quantized and translated to FPGA code via hls4ml. Results on simulated data show four-nines fidelity with ARB enabling adaptation to hardware drift, and a feasible hardware translation path with modest resource usage and latency. The approach highlights practical pathways and limitations for real-device deployment, with broad portability to other quantum architectures that require efficient, near-qubit-control ML solutions.

Abstract

Here we present a technique for using machine learning (ML) for single-qubit gate synthesis on field programmable logic for a superconducting transmon-based quantum computer based on simulated studies. Our approach is multi-stage. We first bootstrap a model based on simulation with access to the full statevector for measuring gate fidelity. We next present an algorithm, named adapted randomized benchmarking (ARB), for fine-tuning the gate on hardware based on measurements of the devices. We also present techniques for deploying the model on programmable devices with care to reduce the required resources. While the techniques here are applied to a transmon-based computer, many of them are portable to other architectures.

Figures (11)

• Figure 1: Total Variation of Pulse Parameters Over Angle. This overlay plot compares the original, averaged, and smoothed datasets, highlighting changes in parameter stability across different processing stages.
• Figure 2: Comparative subplots of output B-spline coefficients 2 to 7 out of 20 coefficients for X gates, demonstrating parallel trends. The six coefficients shown have similar values (y-axis) for a given X gate rotation angle (x-axis), and hence they can be replaced by the average value for data smoothing as explained in the main text.
• Figure 3: Trace Fidelity comparison between a single seed and after dataset optimizations over varying input angles.
• Figure 4: Illustration of the smallest Keras model configuration with 33 parameters achieving four nines of fidelity.
• Figure 5: The first three experiments with artificially perturbed gates (K=500).