Dreaming of Atmospheres
When
Where
Dr. Ingo Waldmann
Postdoctoral Research Fellow
University College London
The field of exoplanetary spectroscopy is as fast moving as it is new. Analysing currently available observations of exoplanetary atmospheres often invoke large and correlated parameter spaces that can be difficult to map or constrain. This is true for both: the data analysis of observations as well as the theoretical modelling of their atmospheres. Issues of low signal-to-noise data and large, non-linear parameter spaces are nothing new and commonly found in many fields of engineering and the physical sciences. Recent years have seen vast improvements in statistical data analysis and machine learning that have revolutionized fields as diverse as telecommunication, pattern recognition, medical physics and cosmology.
In this seminar I will discuss how these improvements in machine learning can be applied to exoplanetary spectroscopy. In particular, how concepts of information theory and statistical entropy can be used to solve some long standing problems in the data analysis and interpretation of exoplanetary atmospheric data. In the first part of this seminar, I will present how the ‘Cocktail Party Problem’ can be used to de-trend Spitzer and Hubble Space Telescope data without any prior knowledge of the instrument and what we can learn from modern ECG heart-rate scanners. In the second part, I will discuss how deep belief neural networks can learn to recognize exoplanetary spectra, provide artificial intelligences to state-of-the-art atmospheric retrieval algorithms and what their ‘dreams’ tell us about atmospheric characteristics. Finally, I will conclude by briefly discussing how we can use information geometry to map the theoretical limits of atmospheric retrievability for current and future space and ground-based instruments.
Host: Dr. Caitlin Griffith