Automated Molecular Discovery with Probabilistic Deep Learning
Abstract
With the advent of wide bandwidth radio-wave receivers, the acquisition of extremely line-rich molecular spectra have become routine in the laboratory and with radio telescopes. With respect to the Submillimeter Array, for example, up to 128 GHz of on-sky frequency coverage will soon be possible. When combined with sub-arcsecond angular resolution and sub-km/s spectral resolution, the physics and chemistry of many astronomical objects or complex mixtures in the laboratory can be analyzed in exquisite molecular detail. However, this new paradigm comes with data collection rates that far exceed our capacity to process and analyze this large body of high-quality data. Recently, we have developed a suite of open-source tools to assist chemists, spectroscopists, and astronomers in analyzing wide bandwidth spectra that emphasize efficiency, automation and reproducibility. In this talk, I will introduce new probabilistic deep learning models we have created to perform spectral assignment, as well as the subsequent identification of unknown molecules in terms of elemental composition and functional groups in a completely automated fashion. Within a Bayesian framework, these models provide a practical means to perform statistical inference on large sets of spectroscopic data, which helps guide spectroscopic analysis with minimal available information.
- Publication:
-
43rd COSPAR Scientific Assembly. Held 28 January - 4 February
- Pub Date:
- January 2021
- Bibcode:
- 2021cosp...43E1998L