Evaluating state-of-the-art cloud quantum computers for quantum neural networks in gravitational waves data analysis

TL;DR

This work assesses the practicality of deploying a quantum neural network for gravitational-wave data analysis on state-of-the-art cloud quantum processors. It documents a hardware-centric evaluation across multiple providers, detailing costs, access delays, and fidelity outcomes, and highlights significant obstacles such as high run costs and frequent Qiskit compatibility issues. The results show high-fidelity performance on shallow circuits but poor predictive accuracy and prohibitive costs for deeper training on current hardware, underscoring that cloud quantum computers are currently best suited for proof-of-concept experiments rather than operational deployment in LISA-era pipelines. The study argues that, in the near term, simulators will remain essential for QML development, while hardware improvements, price-model realignments, and streamlined software ecosystems are critical for practical cloud-based quantum learning in gravitational-wave analysis.

Abstract

In this work, we explore the possibility of using quantum computers provided for usage in cloud by big companies (such as IBM, IonQ, IQM Quantum Computers, etc.) to run our quantum neural network (QNN) developed for data analysis in the context of LISA Space Mission, developed with the Qiskit library in Python. Our previous work demonstrated that our QNN learns patterns in gravitational wave (GW) data much faster than a classical neural network, making it suitable for fast GW signal detection in future LISA data streams. Analyzing the fees from hardware providers like IBM Quantum, Amazon Braket and Microsoft Azure, we found that the fees for running the first segment of our QNN sum up to \$2000, \$60000, and \$1000000 respectively. Using free plans, we succeed to run the 3-qubit feature map of the QNN for one random data sample on {\fontfamily{qcr} \selectfont ibm\_kyoto} and {\fontfamily{qcr}\selectfont IQM Quantum Computers\_Garnet} quantum computers, obtaining a fidelity of 99\%; we could also run the first prediction segment of our QNN on {\fontfamily{qcr} \selectfont ibm\_kyoto}, implemented for 4 qubits, and obtained a prediction accuracy of 20\%. We queried providers such as IBM Quantum, Amazon Braket, Pasqal, and Munich Quantum Valley to obtain access to their plans, but, with the exception of Amazon Braket, our applications remain unanswered to this day. Other major setbacks in using the quantum computers we had access to included Qiskit library version issues (as in the cases of IBM Quantum and IQM Quantum Computers) and the frequent unavailability of the devices, as was the case with the Microsoft Azure provider. All the results presented in this paper were accumulated in 2024.

Figures (4)

• Figure 1: Data samples generation flow for each step. Data samples are generated from a time domain dataset using sliding windows: adjacent windows and overlapping windows for the 1 step and for the 2 step respectively.
• Figure 2: The quantum circuit we ran on ibm_kyoto (left) and the obtained fidelity (right).
• Figure 3: The quantum circuit (part of the sQNN) that was ran on ibm_kyoto for making predictions on 1802 samples of data. The classification accuracy was 18.701%, while on the simulator it was 100%.
• Figure 4: Results obtained when executing the sQNN for prediction making with ibm_kyoto, showing that the learned decision boundary is not preserved and the hardware induces a strong bias towards class 0. Panel (a): Confusion matrix for predictions obtained with ibm_kyoto. Panel (b): Class distribution comparison between the true labels and labels obtained with ibm_kyoto. Panel (c): Accuracy comparison between ideal accuracy and the accuracy obtained with ibm_kyoto. Panel (d): Prediction probability histogram obtained with ibm_kyoto.