Spiking neural networks for streaming qubit readout

Barry M. Dillon; Aqib Javed; Jim Harkin; Patryk Dabkowski; Benjamin Lienhard

quant-phOct 1, 2026

Spiking neural networks for streaming qubit readout

Authors: Barry M. Dillon, Aqib Javed, Jim Harkin, Patryk Dabkowski, Benjamin Lienhard

Organizations: ISRC, Ulster University, Derry, BT48 7JL, Northern Ireland · Technical University of Munich, Department of Physics, 85748 Garching, Germany · Walther-Meißner-Institut, Bayerische Akademie der Wissenschaften, 85748 Garching, Germany · Munich Center for Quantum Science and Technology, 80799 Munich, Germany · Zurich Instruments, 8005 Zürich, Switzerland

Abstract

Fast and accurate qubit-state assignment is essential for feedback, calibration, and error correction in quantum processors. In superconducting platforms, frequency-multiplexed readout makes this task intrinsically multivariate as measured traces can encode crosstalk, qubit-state relaxation events, and other transient nonidealities that are not fully captured by conventional matched filtering. Here, we introduce spiking neural network (SNN) discriminators for superconducting qubit readout. By processing the measurement window in successive time chunks, the networks exploit temporal structure and update classification scores as data arrive, rather than waiting until the end of the readout window. The spiking networks outperform matched-filter discrimination and approach the accuracy of a full-trace artificial neural network. Beyond reaching the performance of artificial neural networks, the key advantage of SNNs is that they provide a streaming, time-resolved estimate of the qubit state that evolves as the readout signal is acquired. Using quantisation-aware training and hls4ml synthesis, we further demonstrate that each FPGA inference update can be completed before the next readout chunk arrives. These results establish spiking neural networks as a promising route to low-latency, real-time qubit readout on FPGA hardware, with broader implications for time-critical quantum-control and scientific-inference applications.

Figures & tables

Figure 1: This figure depicts the real-time updates of the SNN measuring qubits in the

01110

state over

200

ns timesteps. The coloured lines represent the demodulated traces for each qubit. These are then averaged and fed directly to the SNN, which updates its internal neuron states and predictions for the qubit readout classes.

Figure 2: Local

L_{5}

feature values (top) and cumulative

C_{5}

feature values (bottom) for

q_{5}

, comparing

T\!=\!5

and

T\!=\!10

readout summaries over the

1\,\mu\mathrm{s}

acquisition window. Solid lines show the

I

component and dashed lines show the

Q

component.

Hyperparameter	ANN	SNN
Hidden size(s)	$128$	$32,64,128$
Input features	$L_{5},\;L_{5}+C_{5}$	$L_{5},\;L_{5}+C_{5}$
Timesteps	$5,10$	$5,10$
Dropout	$0.02$	$0.02$
Optimiser	AdamW	AdamW
Learning rate	$1.2\times 10^{-3}$	$1.0\times 10^{-2}$

Table 1: Model and training hyperparameters used in training the ANN baseline and the SNN.

Model	Features	$F_{\mathrm{geom}}$	$F_{q1}$	$F_{q2}$	$F_{q3}$	$F_{q4}$	$F_{q5}$
Matched filter	–	0.8962	0.9688	0.7396	0.9007	0.9360	0.9569
ANN, $T\!=\!5$ , $128$ -dim hidden	$L_{5}+C_{5}$	0.9087	0.9698	0.7418	0.9393	0.9459	0.9694
ANN, $T\!=\!10$ , $128$ -dim hidden	$L_{5}$	0.9100	0.9705	0.7435	0.9416	0.9465	0.9704
SNN, $T\!=\!5$ , $64$ -dim hidden	$L_{5}+C_{5}$	0.9070	0.9688	0.7404	0.9372	0.9449	0.9665
SNN, $T\!=\!10$ , $64$ -dim hidden	$L_{5}+C_{5}$	0.9083	0.9695	0.7421	0.9388	0.9453	0.9683
SNN, $T\!=\!10$ , $128$ -dim hidden	$L_{5}+C_{5}$	0.9084	0.9693	0.7424	0.9388	0.9454	0.9685

Table 2: Summary of the readout results for the ANN baseline and the SNN models.

Figure 3: In the plot, we compare ANN models using the geometric mean assignment fidelity. In the plot below, we compare SNN models with a

64

-dim hidden layer using the geometric mean assignment fidelity.

Panel of Figure 4 — Figure 4: Cross-fidelity matrices for MF baseline (top left), the ANN baseline (top right), the $32$ -dim SNN (bottom left), and $64$ -dim SNN (bottom right). All models use $T\!=\!10$ and the $L_{5}+C_{5}$ feature set.

Figure 5: Mid-readout fidelity for the

64

-dim SNN trained on

L_{5}+C_{5}

features. The final-loss-only model is compared with a model trained with equal loss terms at

600\,\mathrm{ns}

and

1\,\mu\mathrm{s}

. The dotted vertical line marks the auxiliary-loss timestep.

Timesteps	Time	Final loss only	$600\,\mathrm{ns}$ + final loss	Change
5	$600\,\mathrm{ns}$	0.8881	0.8920	$+0.0039$
5	$1\,\mu\mathrm{s}$	0.9070	0.9063	$-0.0007$
10	$600\,\mathrm{ns}$	0.8885	0.8939	$+0.0055$
10	$1\,\mu\mathrm{s}$	0.9083	0.9076	$-0.0007$

Table 3: Effect of the

600\,\mathrm{ns}

auxiliary loss on the

64

-dim

L_{5}+C_{5}

SNN.

Figure 6: Mid-readout fidelity for each qubit individually with the

64

-dim SNN.

Figure 7: Test fidelity results for the QAT precision sweep with SNNs with both

64

- and

32

-dim hidden layers, and chunk sizes of

100

ns (

T\!=\!10

Features	Precision	$F_{\mathrm{geom}}$	$F_{q1}$	$F_{q2}$	$F_{q3}$	$F_{q4}$	$F_{q5}$
$L_{5}$	<8,4>	0.9051	0.9680	0.7383	0.9309	0.9442	0.9671
$L_{5}$	<10,4>	0.9061	0.9683	0.7401	0.9327	0.9447	0.9672
$L_{5}$	<12,4>	0.9063	0.9683	0.7411	0.9328	0.9446	0.9672
$L_{5}$	<16,4>	0.9066	0.9684	0.7412	0.9337	0.9448	0.9673
$L_{5}+C_{5}$	<8,4>	0.9066	0.9688	0.7405	0.9370	0.9447	0.9646
$L_{5}+C_{5}$	<10,4>	0.9075	0.9693	0.7414	0.9373	0.9449	0.9673

Table 4: QAT results for the SNN with

T\!=\!10

and a

64

-dim hidden layer.

Hyperparameter	Value
Precision sweep	${W}{4}$ , $W\in\{8,10,12,16\}$
Readout precision	<16,8>
Accumulator precision	<18,8>
Backend	Vitis
FPGA part	xczu49dr-ffvf1760-2-e
I/O interface	io_parallel

Table 5: Hardware hyperparameters used for the FPGA synthesis sweep.

Features	Hidden	Precision	Clk est. (ns)	II	Lat. cyc.	Lat. (ns)	LUT	LUT (%)	FF	FF (%)	DSP	DSP (%)
$L_{5}$	32	$\langle 8,4\rangle$	6.4	1	5	32	22,793	5.4	1,736	0.2	0	0.0
$L_{5}$	32	$\langle 10,4\rangle$	7.7	1	5	39	34,569	8.1	1,960	0.2	0	0.0
$L_{5}$	32	$\langle 12,4\rangle$	7.7	1	4	31	29,251	6.9	1,846	0.2	247	5.8
$L_{5}$	32	$\langle 16,4\rangle$	8.0	1	4	32	29,442	6.9	2,181	0.3	441	10.3
$L_{5}$	64	$\langle 8,4\rangle$	6.5	1	6	39	44,601	10.5	3,142	0.4	0	0.0
$L_{5}$	64	$\langle 10,4\rangle$	7.7	1	5	39	67,527	15.9	3,822	0.4	0	0.0

Table 6: Per-timestep Vitis HLS C-synthesis estimates for the SNN hardware sweep on the XCZU49DR target. All runs use the parameters defined in Tab. 5 . Latency is reported per HLS top-call/chunk.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Representative digitally-unmodulated and demodulated readout trace.

Figure A.2: Average demodulated ground- and excited-state traces for the five qubits.

Explore similar work

Real-Time Detection of Charge Jumps in Superconducting Qubits with a Convolutional Neural Network
Jul 15, 2026Daniel Gaytan-Villarreal, Peter Meiring, Daniel Baxter +6Qubit Convolutional Neural Networks
Department of Physics, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Quantum Division, Fermi National Accelerator Laboratory, Batavia, IL 60510, USA · Department of Physics & Astronomy, Northwestern University, Evanston, IL 60208, USA +1
Stochastic Quantum Spiking Neural Networks with Quantum Memory and Local Learning
Jun 26, 2025Jiechen Chen, Bipin Rajendran, Osvaldo SimeoneQuantum Neural Networks Spiking Neural Networks
Department of Engineering, King’s College London, London, WC2R 2LS, UK · Institute for Intelligent Networked Systems, Northeastern University London, One Portsoken Street, London, E1 8PH, UK
Spiking Neural Network inference on FPGAs with hls4ml
Jun 8, 2026Barry M. DillonSpiking Neural Networks Field-Programmable Gate Arrays
ISRC, Ulster University, Derry, BT48 7JL, Northern Ireland

Jul 15, 2026, quant-phJ/K move · Enter open · S save

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

Daniel Gaytan-Villarreal, Peter Meiring, Daniel Baxter, Daniel Bowring+5

Department of Physics, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Quantum Division, Fermi National Accelerator Laboratory, Batavia, IL 60510, USA · Department of Physics & Astronomy, Northwestern University, Evanston, IL 60208, USA+1

Qubit Convolutional Neural Networks

Open paper AskPDF arXiv

Ionizing radiation from cosmic rays and gammas can induce discontinuous jumps in the environmental charge of superconducting qubits (charge jumps), causing correlated errors that challenge fault-tolerant quantum computing while simultaneously providing a detection signature for quantum sensing applications. Current detection methods operate offline, introducing latency incompatible with in-the-loop qubit control. In this paper, an online detector of charge jumps for superconducting qubits, based on a dilated causal convolutional neural network (DCCNN) designed for in-the-loop deployment on the Quantum Instrumentation Control Kit (QICK) platform, is presented. The network is trained on synthetic Ramsey tomography scans generated from qubit templates measured at the Northwestern Experimental Underground Site (NEXUS) at Fermilab, and translated to FPGA firmware via hls4ml with ap_fixed

\langle 16,6 \rangle

quantization, reaching a per-inference latency of

6.19 μ

s on the Zynq UltraScale+ RFSoC ZCU216. At this operating point the DCCNN matches the detection efficiency of the established offline

χ^2

algorithm (

0.843 \pm 0.022

vs.

0.866 \pm 0.020

|Δq| \in [0.1, 0.5] e

at matched false-positive rate), while requiring no per-qubit hyperparameter tuning. This shifts charge-jump detection from a post-hoc diagnostic to a control-loop primitive, enabling adaptive protocols that respond to radiation-induced events in situ, with applications to quantum-computing error mitigation and to the use of superconducting qubits as particle detectors.

Abstract

Explore similar work

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

Stochastic Quantum Spiking Neural Networks with Quantum Memory and Local Learning

Spiking Neural Network inference on FPGAs with hls4ml