Classifying CMB time-ordered data through deep neural networks. Issue 3 (13th April 2020)

• Record Type: Journal Article
• Authors: Rojas, Felipe; Maurin, Loïc; Dünner, Rolando; Pichara, Karim
• Is Part Of: Monthly notices of the Royal Astronomical Society. Volume 494:Issue 3(2020)
• Page Start: 3741
• Page End: 3749
• Publication Date: 2020-04-13
• Subjects: methods: data analysis -- cosmology: observations -- cosmology: cosmic background radiation; Astronomy -- Periodicals; Periodicals; 520.5
• DOI: 10.1093/mnras/staa1009
• Languages: English
• ISSNs: 0035-8711

Abstract

ABSTRACT: The Cosmic Microwave Background (CMB) has been measured over a wide range of multipoles. Experiments with arcminute resolution like the Atacama Cosmology Telescope (ACT) have contributed to the measurement of primary and secondary anisotropies, leading to remarkable scientific discoveries. Such findings require careful data selection in order to remove poorly behaved detectors and unwanted contaminants. The current data classification methodology used by ACT relies on several statistical parameters that are assessed and fine-tuned by an expert. This method is highly time-consuming and band or season-specific, which makes it less scalable and efficient for future CMB experiments. In this work, we propose a supervised machine learning model to classify detectors of CMB experiments. The model corresponds to a deep convolutional neural network. We tested our method on real ACT data, using the 2008 season, 148 GHz, as training set with labels provided by the ACT data selection software. The model learns to classify time-streams starting directly from the raw data. For the season and frequency considered during the training, we find that our classifier reaches a precision of 99.8 per cent. For 220 and 280 GHz data, season 2008, we obtained 99.4 per cent and 97.5 per cent of precision, respectively. Finally, we performed a cross-season test over 148 GHz data from 2009 and 2010 for which our model reaches a precision of 99.8 per cent and 99.5 per cent, respectively. Our model is about 10x faster than the current pipeline, making it potentially suitable for real-time implementations.