Figure 1
The results and some snapshots of our project.
Courtesy of Jaewon Jung.
Figure 2
The simple sketch of how our model works.
Courtesy of Jaehwan Kim who led this project.
Figure 3
The simple sketch of how our model learn environment model.
Courtesy of Jaehwan Kim who led this project.
Figure 4
The simple sketch of how our model learn signal model.
Courtesy of Jaehwan Kim who led this project.
Contributions: J. Jung ran a hardware and numerical experiments and analyzed the results. J. Jung developed a protocol related to the use of IBM QPUs. J. Kim and S. Ryu developed machine learning protocols and classical simulation methods.
Machine Learning Jun. 2021 – Jun. 2022
Optimizing Microwave Pulses for Arbitrary Unitary Gates via Neural ODEs
⌜Under the IBM Researchers program which is an exclusive program for researchers experienced in the quantum information area (now an IBM Quantum Credits program), our team built a neural network that can learn both environment and signal Hamiltonian and
constructed pulse-defined arbitrary unitary gate both in simulation and IBM’s real backend using Qiskit and Torch.⌟
Ordinary differential equations are fundamental differential equations that describe the dynamics of nature. We utilized the newly built "Neural Ordinary Differential Equations" at the moment to construct the optimized pulse-defined gates for quantum computers.
Our work is to construct and optimize unitary gates on a few qubits (~10 qubits) into one pulse-defined quantum gate. Figs 2-4 show holistic explanations of our methods. For n-qubit quantum operations over real quantum hardware, 1. we only learn "env" parameters over l-qubits for all possible decomposable Hamiltonians then 2. learn "signal" parameters. Therefore, we can reduce the learning time complexity of env models (see terminology in Fig. 2). Our work is first tested on a classical simulation and then tested on real quantum hardware using IBM Quantum's QPUs utilizing sinusoidal representation networks (SIRENs), wavelet transforms and learning strategy optimizations with rigorous hyperparameter tuning and showed high accuracy in learning both environment and signal parameters for arbitrary unitary gates as seen in Fig. 1.