Extraction of the redshifted HI global signal using quantum neural networks

Abstract

The primary goal of ongoing and upcoming experiments, such as EDGES, SARAS, and SKAO, is to explore and analyze key epochs in the universe, including the Dark Ages, Cosmic Dawn, and Epoch of Reionization. These experiments utilize single-dish antennae radio telescopes to measure the sky-averaged differential brightness temperature of the redshifted HI 21cm signal, resulting in the Global signal represented as a function of redshift. However, a significant challenge in these observations lies in the dominating nature of the foregrounds in the radio band. Additionally, Earth’s ionosphere and Radio Frequency Interference introduce further complications in the observation process. Consequently, accurate detection of the desired cosmological signal requires effective corruption detection and removal. In this work, we propose the use of Quantum Neural Networks (QNN) as an alternative deep learning technique to address foreground and thermal noise contamination and extract the redshifted global HI signal buried under these corruptions. This work introduces a novel methodology using QNNs that compares to the prediction estimation of ANNs while also reducing the computational time the quantum simulation takes. Feature extraction/reduction methods like kernel PCA provide a better representation of the features for the quantum network and help reduce the complexity of the network, thereby avoiding barren plateaus (vanishing gradients). Our results indicate that the methodology proposed successfully identifies signal parameters without (with) any (constant) contamination with a test R2 score ! 0.98. However, the model encounters challenges in recovering the signal in the presence of varying foreground and thermal noise, predicting only the foreground parameters up to the first order due to the dominance of a0 and a1 foreground parameters. To remove the effects of foreground, a prior foreground removal technique based on PCA was implemented before the network training, leading to better training and prediction with a test R2 score " 0.85. Notably, the proposed pipeline for parameter estimation using QNNs showcases a computationally efficient and fast quantum simulation by utilizing GPU and Just-In-Time (jit) compilation while also maintaining a good model prediction in comparison to a hybrid quantum-classical neural network employed alongside the quantum model proposed in the pipeline.