اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدPeeking inside the Black Box: Interpreting Deep Learning Models for Exoplanet Atmospheric Retrievals
311
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Deep learning algorithms are growing in popularity in the field of exoplanetary science due to their ability to model highly non-linear relations and solve interesting problems in a data-driven manner. Several works have attempted to perform fast retrievals of atmospheric parameters with the use of machine learning algorithms like deep neural networks (DNNs). Yet, despite their high predictive power, DNNs are also infamous for being black boxes. It is their apparent lack of explainability that makes the astrophysics community reluctant to adopt them. What are their predictions based on? How confident should we be in them? When are they wrong and how wrong can they be? In this work, we present a number of general evaluation methodologies that can be applied to any trained model and answer questions like these. In particular, we train three different popular DNN architectures to retrieve atmospheric parameters from exoplanet spectra and show that all three achieve good predictive performance. We then present an extensive analysis of the predictions of DNNs, which can inform us - among other things - of the credibility limits for atmospheric parameters for a given instrument and model. Finally, we perform a perturbation-based sensitivity analysis to identify to which features of the spectrum the outcome of the retrieval is most sensitive. We conclude that for different molecules, the wavelength ranges to which the DNNs predictions are most sensitive, indeed coincide with their characteristic absorption regions. The methodologies presented in this work help to improve the evaluation of DNNs and to grant interpretability to their predictions.
قيم البحث
اقرأ أيضاً
Over the past decade, the study of extrasolar planets has evolved rapidly from plain detection and identification to comprehensive categorization and characterization of exoplanet systems and their atmospheres. Atmospheric retrieval, the inverse mode
ling technique used to determine an exoplanetary atmospheres temperature structure and composition from an observed spectrum, is both time-consuming and compute-intensive, requiring complex algorithms that compare thousands to millions of atmospheric models to the observational data to find the most probable values and associated uncertainties for each model parameter. For rocky, terrestrial planets, the retrieved atmospheric composition can give insight into the surface fluxes of gaseous species necessary to maintain the stability of that atmosphere, which may in turn provide insight into the geological and/or biological processes active on the planet. These atmospheres contain many molecules, some of them biosignatures, spectral fingerprints indicative of biological activity, which will become observable with the next generation of telescopes. Runtimes of traditional retrieval models scale with the number of model parameters, so as more molecular species are considered, runtimes can become prohibitively long. Recent advances in machine learning (ML) and computer vision offer new ways to reduce the time to perform a retrieval by orders of magnitude, given a sufficient data set to train with. Here we present an ML-based retrieval framework called Intelligent exoplaNet Atmospheric RetrievAl (INARA) that consists of a Bayesian deep learning model for retrieval and a data set of 3,000,000 synthetic rocky exoplanetary spectra generated using the NASA Planetary Spectrum Generator. Our work represents the first ML retrieval model for rocky, terrestrial exoplanets and the first synthetic data set of terrestrial spectra generated at this scale.
We demonstrate the effectiveness of the Exoplanet Characterisation Observatory mission concept for constraining the atmospheric properties of hot and warm gas giants and super Earths. Synthetic primary and secondary transit spectra for a range of pla
nets are passed through EChOSim (Waldmann & Pascale 2014) to obtain the expected level of noise for different observational scenarios; these are then used as inputs for the NEMESIS atmospheric retrieval code and the retrieved atmospheric properties (temperature structure, composition and cloud properties) compared with the known input values, following the method of Barstow et al. (2013a). To correctly retrieve the temperature structure and composition of the atmosphere to within 2 {sigma}, we find that we require: a single transit or eclipse of a hot Jupiter orbiting a sun-like (G2) star at 35 pc to constrain the terminator and dayside atmospheres; 20 transits or eclipses of a warm Jupiter orbiting a similar star; 10 transits/eclipses of a hot Neptune orbiting an M dwarf at 6 pc; and 30 transits or eclipses of a GJ1214b-like planet.
Future space-based direct imaging missions will perform low-resolution (R$<$100) optical (0.3-1~$mu$m) spectroscopy of planets, thus enabling reflected spectroscopy of cool giants. Reflected light spectroscopy is encoded with rich information about t
he scattering and absorbing properties of planet atmospheres. Given the diversity of clouds and hazes expected in exoplanets, it is imperative we solidify the methodology to accurately and precisely retrieve these scattering and absorbing properties that are agnostic to cloud species. In particular, we focus on determining how different cloud parameterizations affect resultant inferences of both cloud and atmospheric composition. We simulate mock observations of the reflected spectra from three top priority direct imaging cool giant targets with different effective temperatures, ranging from 135 K to 533 K. We perform retrievals of cloud structure and molecular abundances on these three planets using four different parameterizations, each with increasing levels of cloud complexity. We find that the retrieved atmospheric and scattering properties strongly depend on the choice of cloud parameterization. For example, parameterizations that are too simplistic tend to overestimate the abundances. Overall, we are unable to retrieve precise/accurate gravity beyond $pm$50%. Lastly, we find that even low SNR=5, low R=40 reflected light spectroscopy gives cursory zeroth order insights into cloud deck position relative to molecular and Rayleigh optical depth level.
We introduce a new machine learning based technique to detect exoplanets using the transit method. Machine learning and deep learning techniques have proven to be broadly applicable in various scientific research areas. We aim to exploit some of thes
e methods to improve the conventional algorithm based approaches presently used in astrophysics to detect exoplanets. Using the time-series analysis library TSFresh to analyse light curves, we extracted 789 features from each curve, which capture the information about the characteristics of a light curve. We then used these features to train a gradient boosting classifier using the machine learning tool lightgbm. This approach was tested on simulated data, which showed that is more effective than the conventional box least squares fitting (BLS) method. We further found that our method produced comparable results to existing state-of-the-art deep learning models, while being much more computationally efficient and without needing folded and secondary views of the light curves. For Kepler data, the method is able to predict a planet with an AUC of 0.948, so that 94.8 per cent of the true planet signals are ranked higher than non-planet signals. The resulting recall is 0.96, so that 96 per cent of real planets are classified as planets. For the Transiting Exoplanet Survey Satellite (TESS) data, we found our method can classify light curves with an accuracy of 0.98, and is able to identify planets with a recall of 0.82 at a precision of 0.63.
In an effort to interpret black-box models, researches for developing explanation methods have proceeded in recent years. Most studies have tried to identify input pixels that are crucial to the prediction of a classifier. While this approach is mean
ingful to analyse the characteristic of blackbox models, it is also important to investigate pixels that interfere with the prediction. To tackle this issue, in this paper, we propose an explanation method that visualizes undesirable regions to classify an image as a target class. To be specific, we divide the concept of undesirable regions into two terms: (1) factors for a target class, which hinder that black-box models identify intrinsic characteristics of a target class and (2) factors for non-target classes that are important regions for an image to be classified as other classes. We visualize such undesirable regions on heatmaps to qualitatively validate the proposed method. Furthermore, we present an evaluation metric to provide quantitative results on ImageNet.
الأسئلة المقترحة
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
