<p>Detecting the redshifted H<span>i</span> 21 cm signal (both emission and absorption feature against the Cosmic Microwave Background (CMB)) is a key goal of cosmological experiments, offering insights into the Cosmic Dawn and Epoch of Reionization. However, strong foregrounds, ionospheric effects, and Radio Frequency Interference (RFI) pose significant challenges. This work explores the first utilization of quantum neural networks (QNNs) as an alternative deep-learning approach for extracting the global 21 cm signal parameters from these contaminations and comparing their efficacy against classical Neural Networks. We utilize the Accelerated Reionization Era Simulations (ARES) code to simulate global signals varying three target astrophysical parameters. We develop a hybrid quantum-classical framework to estimate these parameters from synthetic datasets, both with and without foregrounds and thermal noise. To mitigate foreground contamination, we incorporate a Principal Component Analysis (PCA)-based removal technique. Additionally, Graphics Processing Unit (GPU) and Just-In-Time (JIT) compilation are used to enhance computational efficiency. In the absence of contamination, the QNN achieves a test <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\mathrm R^{2}\)</EquationSource> </InlineEquation> score of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\ge \)</EquationSource> </InlineEquation> 0.98 for parameter estimation. For contaminated signals, the architecture initially struggles due to the dominance of lowest-order polynomial terms of the foreground model (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\mathrm a_{0}\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\mathrm a_{1}\)</EquationSource> </InlineEquation>). However, implementing PCA-based foreground removal improves model performance, yielding a test <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\mathrm{R}^{2}\)</EquationSource> </InlineEquation> score of <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\approx \)</EquationSource> </InlineEquation> 0.85. This study provides an initial benchmark of QNNs for global 21-cm signal estimation. QNNs match classical performance for signal-only data but struggle with foreground-contaminated spectra, where PCA improves results. Our hybrid quantum-classical framework highlights both current limitations and future opportunities for circuit design, hyperparameter tuning, and contamination mitigation in cosmological applications.</p>

Can Quantum Neural Networks Estimate the Redshifted Global Hi 21-cm Signal Parameters?

期刊论文
发表日期：

2026年3月21日

查看全文

Akash Gowtham,
Anshuman Tripathi,
Abhirup Datta

摘要

Detecting the redshifted Hi 21 cm signal (both emission and absorption feature against the Cosmic Microwave Background (CMB)) is a key goal of cosmological experiments, offering insights into the Cosmic Dawn and Epoch of Reionization. However, strong foregrounds, ionospheric effects, and Radio Frequency Interference (RFI) pose significant challenges. This work explores the first utilization of quantum neural networks (QNNs) as an alternative deep-learning approach for extracting the global 21 cm signal parameters from these contaminations and comparing their efficacy against classical Neural Networks. We utilize the Accelerated Reionization Era Simulations (ARES) code to simulate global signals varying three target astrophysical parameters. We develop a hybrid quantum-classical framework to estimate these parameters from synthetic datasets, both with and without foregrounds and thermal noise. To mitigate foreground contamination, we incorporate a Principal Component Analysis (PCA)-based removal technique. Additionally, Graphics Processing Unit (GPU) and Just-In-Time (JIT) compilation are used to enhance computational efficiency. In the absence of contamination, the QNN achieves a test \(\mathrm R^{2}\) score of \(\ge \) 0.98 for parameter estimation. For contaminated signals, the architecture initially struggles due to the dominance of lowest-order polynomial terms of the foreground model ( \(\mathrm a_{0}\) , \(\mathrm a_{1}\) ). However, implementing PCA-based foreground removal improves model performance, yielding a test \(\mathrm{R}^{2}\) score of \(\approx \) 0.85. This study provides an initial benchmark of QNNs for global 21-cm signal estimation. QNNs match classical performance for signal-only data but struggle with foreground-contaminated spectra, where PCA improves results. Our hybrid quantum-classical framework highlights both current limitations and future opportunities for circuit design, hyperparameter tuning, and contamination mitigation in cosmological applications.