2021-10-26, 10:15–10:30, Grand Ballroom
Astronomical spectra can be made up of hundreds or even thousands of emission and absorption lines. Astronomers need to identify each line in order to be able to determine the physical conditions of the objects studied as the temperature of the sources and the column densities of the observed species (molecules and/or atoms). The constant improvement of instruments in terms of sensitivity, instantaneous spectral range, and spatial resolution capabilities produces a mass of broad spectra with high spectral resolution for which handmade line identification is ineffective and even maybe impossible.
With the advent of BIG DATA, AI algorithms have proven to be very effective in solving complex problems (mainly related to classification and prediction tasks) for many different fields including astrophysics. The aim of this study is to automate the identification of species from their observed lines in rich astronomical spectra by combining methods in signal processing and machine learning with expert knowledge.
This talk will cover three solutions based on (1) wavelets transform, expert knowledge and decision trees to identify the species associated to each spectral line, (2) Artificial Neural Networks to predict if a species is present in a spectrum or not and (3) a greedy algorithm that simulate successively the presence of each species of the database (and its isotopes) in order to check its correspondance in the spectrum. Last, we combine and compare these methods to improve our results.
The results of our research, using an ALMA spectrum very rich in molecular lines combined with the use of CDMS and JPL molecular spectroscopic databases, have already allowed us to find a molecule that had never been detected in the spectrum experimented.
Big data: How to deal with the 5 Vs (volume, velocity, variety, veracity, value)