Exoplanet Atmosphere Characterization via Transit Spectra Classification

1. Introduction

As of February 2023, around 5040 exoplanets have been identified [1] but the characterization of exoplanet atmosphere is still a challenging task amid significant progress in improvement of telescopes and data analytics [2]. Only a handful of observed exoplanets have been characterized in terms of atmosphere composition [3], which is vital for various purposes including the search of habitable planets.

Among the various methods of exoplanet observation for atmosphere characterization, transmission spectroscopy is a common method used [4] specially when the orbit of the exoplanet aligns between the host star and the observing telescope (on Earth or around Earth). Typically, the data analysis task to be done given that an actual measurement has been made is by using calibration spectra to estimate from a measured transmission (absorption) spectra the amounts of the atmospheric chemicals. The data analytics approach study presented in this paper reformulates the data analysis problem by looking at it as a classification problem (Figure 1).

Posing the data analytics task as a classification problem requires two data components when training a classifier: (1) spectral data as input, and (2) label of the spectral data as output. This dataset structure can be created using known models about exoplanet atmosphere composition, and one platform that has been developed for simulation of transmission spectra of transiting exoplanet is the PLATON 5.3 [5], which is a Python-based module. With this data analytics workflow of simulating transmission spectra from randomized levels of gases (CO₂, O₂, N₂ and CH₄) via the PLATON 5.3 platform followed by the training of a forward neural network (NN) classifier (Figure 1), we show in this paper the potential of classification algorithm in characterizing the atmosphere composition of transiting exoplanet by using transit depth-versus-wavelength datasets.

2. Methodology

A schematic overview of the data analysis workflow is shown in Figure 1. This data analytics workflow leverages on the capability of transit-depth-versus-wavelength spectral data to capture the characteristic atmosphere fingerprint of the transiting exoplanet [4]. The datasets and Python codes in Jupyter Notebook files used in this study are provided online via the GitHub repository of the work [6].

The transit spectra datasets were simulated using the Python-based module PLATON 5.3 [5] (Figure 1a). The 3-gas mix combinations used in the simulation of spectra are summarized in Table 1. For the atmosphere composition simulations, the following star-planet parameters were used: an exoplanet with size R_p = 1.40*R_jupiter, and mass M_p = 0.73*M_jupiter orbiting around the host star of size M_s = 1.16*M_sun and surface temperature of 1200 Kelvin. These parameter levels were the default settings in the PLATON 5.3 modules example simulation codes [5] and were kept the same in the work. The number of random simulations (n) in each class was varied at n = 10 and n = 100. A sample graphical rendering of the transit depth-versus-wavelength for the spectra Set I is shown in Figure 2.

3. Results

The performance of the trained forward neural network classifiers are shown as follows: Figure 3 for the confusion matrix, Figure 4 for the receiver operating characteristic (ROC) curve, and Table 2 for the summary of precision, recall, and F1-score for each class.

4. Discussion

Overall, the trained classifiers can achieve very good classification performance reaching 97.9% overall accuracy. The higher number of spectral data, which is n = 100 in this study, favors higher prediction accuracy. The levels of precision, recall, and F1-score of the trained classifiers also indicate low misclassification rates. Based on these results of training forward NN classifiers on the transit spectra generated via PLATON 5.3, we conclude that a classification algorithm can be a potential method of characterizing the atmosphere of transiting exoplanets.