Quantum Computing Tools for Fast Detection of Gravitational Waves in the Context of LISA Space Mission

TL;DR

This work develops a low-latency GW-detection tool for the LISA mission using a Variational Quantum Classifier (VQC) with a 4-qubit input. It introduces a two-stage warm-start training regime and window-based data sampling to classify time-series data into GW/noise and merger/non-merger categories, using a Pauli-Z feature map and a Pauli Two Design ansatz totaling 64 trainable parameters. On the Sangria LISA Data Challenge, the QNN achieves over $98\%$ accuracy in identifying noisy GW samples on the blind set and detects $5$ of $6$ mergers with a threshold of $0.44$, albeit missing the lowest-amplitude event, highlighting both the data efficiency and current sensitivity limits relative to a classical RNN baseline that detects all mergers. Compared to the classical approach, the QNN uses dramatically fewer training samples and parameters, offering a promising path toward efficient, hardware-friendly GW detection, with future work aimed at improving sensitivity to weak signals and extending capabilities to parameter estimation and hardware benchmarking.

Abstract

The field of gravitational wave (GW) detection is progressing rapidly, with several next-generation observatories on the horizon, including LISA. GW data is challenging to analyze due to highly variable signals shaped by source properties and the presence of complex noise. These factors emphasize the need for robust, advanced analysis tools. In this context, we have initiated the development of a low-latency GW detection pipeline based on quantum neural networks (QNNs). Previously, we demonstrated that QNNs can recognize GWs simulated using post-Newtonian approximations in the Newtonian limit. We then extended this work using data from the LISA Consortium, training QNNs to distinguish between noisy GW signals and pure noise. Currently, we are evaluating performance on the Sangria LISA Data Challenge dataset and comparing it against classical methods. Our results show that QNNs can reliably distinguish GW signals embedded in noise, achieving classification accuracies above 98\%. Notably, our QNN identified 5 out of 6 mergers in the Sangria blind dataset. The remaining merger, characterized by the lowest amplitude, highlights an area for future improvement in model sensitivity. This can potentially be addressed using additional mock training datasets, which we are preparing, and by testing different QNN architectures and ansatzes.

Figures (17)

• Figure 1: The training dataset for the Sangria Data Challenge, showing both the GW noisy signal and the GW noiseless signal (upper), the blind dataset for the Sangria Data Challenge (middle), and the noise of the Sangria training dataset, obtained by subtracting the clean signal from the noisy signal (lower). 'TDI' stands for Time Delay Interferometry, while 'X' represents the X projection of the signal
• Figure 2: Power spectral density mean versus standard deviation (left) and spectral entropy versus maximum (right) for the noisy waveforms samples, respectively the noise samples in Sangria train and blind datasets. The samples were obtained using the procedure presented in Sec.\ref{['sec:adjacentwindowsapproach']}
• Figure 3: Assigned labels for each individual data point in the Sangria training dataset. This point-wise label list serves as the basis for assigning labels to samples generated via the moving window procedure.
• Figure 4: Final sample labels for the Sangria training dataset, used for the second training stage.
• Figure 5: The feature map of the QNN: 1st order Pauli expansion circuit paulifeaturemap. $x$ is the sample feature vector - classical input - that needs to be encoded in a quantum state with which the quantum program operates. Each qubit contains the information of one vector element $x[i], i=\overline{0,3}$.