Real-Time Detection of Charge Jumps in Superconducting Qubits with a Convolutional Neural Network

Superconducting qubits are highly sensitive to ionizing radiation, which can cause discrete 'charge jumps' that shift the qubit's transition frequency and degrade coherence. These jumps are a significant hurdle for fault-tolerant quantum computing, as they introduce correlated errors that violate standard error-correction assumptions. Traditionally, detecting these events requires offline analysis of Ramsey tomography data, which introduces significant latency and prevents real-time intervention.

The researchers designed a Dilated Causal Convolutional Neural Network (DCCNN) to identify charge jumps in real-time. Unlike standard CNNs, this architecture uses causal convolutions to prevent information leakage from future time steps and dilated convolutions to expand the receptive field without increasing the parameter count. The model was trained on synthetic Ramsey scans and optimized for hardware deployment using Quantization-Aware Training (QAT) to ensure compatibility with FPGA firmware via the hls4ml framework.

The DCCNN achieved a per-inference latency of 6.19 microseconds on a Zynq UltraScale+ RFSoC platform, making it suitable for in-the-loop control. In comparative testing, the DCCNN matched the detection efficiency of the established offline chi-squared algorithm (0.843 vs 0.866) while eliminating the need for per-qubit hyperparameter tuning. This capability allows for adaptive quantum control protocols that can respond to radiation-induced events as they occur.

By moving charge-jump detection from a post-hoc diagnostic to a real-time control primitive, this work enables active error mitigation in quantum processors. Furthermore, it enhances the utility of superconducting qubits as sensitive particle detectors, allowing them to act as in-situ sensors for fundamental physics experiments that require rapid event triggering.