FPGA-based In-situ Learning for Real-time Quantum State Discrimination on QubiCML

Quantum computing systems often face the challenge of qubit readout drift, where qubit shift over time, leading to potential degradation in the fidelity of quantum algorithms. To address this, we propose an enhanced Field-Programmable Gate Array (FPGA) -based quantum state discrimination approach incorporating online learning capabilities directly on the FPGA hardware. By enabling the neural network model to adapt its parameters continuously during inference—using the feedback derived from QUantum BIt Control with Machine Learning’s (QubiCML) real-time measurements and ground truth data—our system can effectively compensate for these drifts in qubit properties. This on-the-fly adaptation ensures that the model remains tuned to the specific operating conditions of the quantum device, thereby maintaining high-fidelity performance without significant manual intervention. Unlike traditional approaches that involve tuning model parameters on host computers or Graphics Processing Units (GPUs), our FPGA-based online learning method drastically reduces the time and resource overhead by integrating the learning process directly on the FPGA, providing real-time updates and optimized performance. This enables users to continuously optimize quantum algorithms for changing conditions, ensuring consistent accuracy and improved robustness in quantum state discrimination.