Pulse Level Compilation of Parameterized Gates via Neural Networks
Session Number
Project ID: PHYS 25
Advisor(s)
Andrew Goldschmidt, University of Chicago
Discipline
Physical Science
Start Date
17-4-2024 10:00 AM
End Date
17-4-2024 10:15 AM
Abstract
Recently, a class of hybrid quantum-classical algorithms called Variational Quantum Algorithms (VQAs) have been investigated as a promising candidate for practical near term algorithms. VQAs require quantum circuits that consist of parameterized gates, which can then be fine tuned to minimize cost functions. In practice, implementing a quantum gate requires compiling it to a sequence of pulses that can be executed on hardware. However, constructing a new set of pulses for each distinct angle encountered in the optimization loop of a VQA is an impractical approach due to the large classical computation overhead required to do so. In this work, we explore a Neural Network solution to the problem. The network is first pre-trained on pulses obtained by using Quantum Collocation (implemented using the PICO framework) for a set of sampled parameter values, and then trained over all parameter values to minimize gate infidelity. We demonstrate that this neural network approach successfully interpolates between provided pre-trained sample pulses, creating high fidelity gates, while also successfully reducing the pulse schedule time for circuits when compared to the standard basis set decompositions. The latter is important for near term applications, where the effects of finite depolarizing and dephasing times are important considerations.
Pulse Level Compilation of Parameterized Gates via Neural Networks
Recently, a class of hybrid quantum-classical algorithms called Variational Quantum Algorithms (VQAs) have been investigated as a promising candidate for practical near term algorithms. VQAs require quantum circuits that consist of parameterized gates, which can then be fine tuned to minimize cost functions. In practice, implementing a quantum gate requires compiling it to a sequence of pulses that can be executed on hardware. However, constructing a new set of pulses for each distinct angle encountered in the optimization loop of a VQA is an impractical approach due to the large classical computation overhead required to do so. In this work, we explore a Neural Network solution to the problem. The network is first pre-trained on pulses obtained by using Quantum Collocation (implemented using the PICO framework) for a set of sampled parameter values, and then trained over all parameter values to minimize gate infidelity. We demonstrate that this neural network approach successfully interpolates between provided pre-trained sample pulses, creating high fidelity gates, while also successfully reducing the pulse schedule time for circuits when compared to the standard basis set decompositions. The latter is important for near term applications, where the effects of finite depolarizing and dephasing times are important considerations.