Towards the prediction of molecular parameters from astronomical emission lines using Neural Networks

• Published in: Experimental astronomy Vol. 52; no. 1-2; pp. 157 - 182
• Main Authors: Barrientos, Alejandro, Holdship, Jonathan, Solar, Mauricio, Martín, Sergio, Rivilla, Víctor M., Viti, Serena, Mangum, Jeff, Harada, Nanase, Sakamoto, Kazushi, Muller, Sébastien, Tanaka, Kunihiko, Yoshimura, Yuki, Nakanishi, Kouichiro, Herrero-Illana, Rubén, Mühle, Stefanie, Aladro, Rebeca, Aalto, Susanne, Henkel, Christian, Humire, Pedro
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.10.2021; Springer Nature B.V
• Subjects: ALCHEMI; Algorithms; Arrays; Astronomy; Carbon monoxide; Chemistry and Earth Sciences; Computer Science; Cubes; Density; Excitation; Machine learning; MADCUBA; Molecular astronomy; Molecular parameters; Neural networks; Noise levels; Noise prediction; Observations and Techniques; Observatories; Original Article; Parameters; Performance prediction; Physics; Physics and Astronomy; Radio telescopes; Spectra; Spectral resolution; Spectrum analysis; Statistics for Engineering; MADCUBA; Molecular parameters; Neural networks; Molecular astronomy; ALCHEMI; Machine learning
• ISSN: 0922-6435; 1572-9508; 1572-9508
• DOI: 10.1007/s10686-021-09786-w

Molecular astronomy is a field that is blooming in the era of large observatories such as the Atacama Large Millimeter/Submillimeter Array (ALMA). With modern, sensitive, and high spectral resolution radio telescopes like ALMA and the Square Kilometer Array, the size of the data cubes is rapidly escalating, generating a need for powerful automatic analysis tools. This work introduces MolPred , a pilot study to perform predictions of molecular parameters such as excitation temperature (T ex ) and column density ( l o g ( N )) from input spectra by the use of neural networks. We used as test cases the spectra of CO, HCO + , SiO and CH 3 CN between 80 and 400 GHz. Training spectra were generated with MADCUBA, a state-of-the-art spectral analysis tool. Our algorithm was designed to allow the generation of predictions for multiple molecules in parallel. Using neural networks, we can predict the column density and excitation temperature of these molecules with a mean absolute error of 8.5% for CO, 4.1% for HCO + , 1.5% for SiO and 1.6% for CH 3 CN. The prediction accuracy depends on the noise level, line saturation, and number of transitions. We performed predictions upon real ALMA data. The values predicted by our neural network for this real data differ by 13% from the MADCUBA values on average. Current limitations of our tool include not considering linewidth, source size, multiple velocity components, and line blending.