GSoC 22 - Finding Exoplanets with Astronomical Observations

00 - SECTION 0 - readme.md

GSoC + ML4SCI [2022] - Finding Exoplanets with Astronomical Observations

01 - SECTION 1 - readme.md

SECTION 1 - CONTEXT

Overview

This Github Gist contains all the code, information and useful links relating to a Google Summer of Code (GSoC) 2022 project, the project is called "Finding Exoplanets with Astronomical Observations".

Project Description

The purpose of this project is to use publicly available data from observations intended to identify exoplanets in order to determine whether or not a given stellar system contains one or more exoplanets.

An exoplanet is any planet not in our solar system. Since the early 1990s, thousands of exoplanets have been discovered. Exoplanets are identified through a variety of methods, including, but not limited to, transit light curve analysis, direct imaging, and radial velocity measurements. One of the most successful attempts was the Kepler Mission, which studied light curves of hundreds of thousands of stars and identified over 2,500 planets. Each detection technique is biased towards detecting different types of planets at different distances from stars, so a variety of methods is necessary to fully explore the exoplanetary parameter space.

As is exemplified by the Kepler Mission, identifying exoplanets – using any method – relies on massive datasets and signals that can be subtle and noisy. In fact, the first exoplanet was observed many years before it was identified because the researchers did not originally notice it in the data. An unfortunate side effect of this delay was that the “first” discovered exoplanet was announced in the interim despite being observed years later. Machine learning may be a useful avenue for systematically, accurately, and quickly identifying planets and avoiding such situations.

Project Outcomes

Using publicly available observational data, models will be constructed that predict the presence of a planet. This may use data from any type of detection of method, and there is no limit to the number of detection methods that may be used.

02 - SECTION 2 - readme.md

SECTION 2 - DATA

Data Sources

There are 3 main types of astronomical observation data, all of which are taken from space telescopes such as NASA's Kepler and TESS missions; these include: target pixel files, lightcurves and stellar information in the form of tabular data. This project used lightcurves as the data is almost in its raw form as extracted from the telescopes and it is in a form that can be loaded into machine learning pipelines with some relatively light processing.

The data was taken and combined from 2 sources: Kepler Object of Interest (KOI) Activity Tables - Cumulative Table from which more information is available here. This returns a list of all the stars that the Kepler mission has observed, but not the lightcurves themselves. To obtain the lightcurves themselves, the python package Lightkurve is used.

03 - Kepler Cumulative Table - readme.md

Kepler Cumulative Table

The following csv file displays the first 5 rows obtained from the NASA Kepler Object of Interest Cumulative Table.

04 - Kepler Cumulative Table.csv

05 - Introduction to Lightkurve Notebook.ipynb

07 - Data Collection Example Notebook.ipynb

08 - SECTION 3 - readme.md

SECTION 3 - DATA PROCESSING

Data Processing Necessity

As seen above, the lightcurves are time-series data with light flux being the variable which evolves over time. This data is not readily inputtable into machine learning models for several reasons: most models apart from a few primitive ones (such as K-Nearest Neighbours) cannot deal with time-series data that is not of equal length; and the same goes for the NaNs present in the data. Additionally, there are long-term trends arising from dying stars and other phenomena which may affect the readings, as well as outliers in the data. All of this means processing the data is necessary.

Processing Methodolgy

To end up with correctly formatted data, 2 approaches were undertaken.

1. Global View: This meant following the approach outlined in this paper to generate a global view: https://arxiv.org/pdf/1712.05044.pdf, which was implemented in python with the Lighkurve package based on this tutorial: https://docs.lightkurve.org/tutorials/3-science-examples/exoplanets-machine-learning-preprocessing.html.
2. Raw View: The idea behind this approach is to take the raw lighcurves, and apply the lowest levels of processing that allow it to be inputted into machine learning models. This involves removing outliers and converting the curves into equally spaced time-steps.

09 - Global View Example Notebook.ipynb

10 - Full Global View Data Creation Notebook.ipynb

11 - Global View Data - readme.md

Global View Data

The following csv file displays the first 5 rows obtained from the above notebook.

12 - Global View Data.csv

14 - Full Raw View Data Creation Notebook.ipynb

15 - Raw View Data - readme.md

Raw View Data

The following csv file displays the first 5 rows obtained from the above notebook.

16 - Raw View Data.csv

17 - SECTION 4 - readme.md

SECTION 4 - MACHINE LEARNING MODELS

Model Sources

In order to be able to apply many models to the data, pre-built models were selected; this came with the added benefits that they were known to be effective for the task of time-series classification, and and performance evaluation and inference was in-built in most of them. These came from 2 sources: the python package tsai and public models from Kaggle

TSAI

Using tsai, we performed a benchmark test (only validation accuracy was recorded) with a variety of models on both datasets, this then guided which models to apply to the datasets; this was done as the time it takes to fully train all these models would have been computationally significant.

Full models:

• ResNet
• XceptionTime
• InceptionTime
• Time Series Transformer
• MiniRocket

Kaggle

The following notebook models were used:

• Exoplanet Hunting: TOP Score using SMOTE and CNN
• Exoplanet detection using Spiking Neural Networks

18 - Model Benchmark Notebook.ipynb

19 - Model Benchmark Results.csv

20 - Global View - TST Model Notebook.ipynb

21 - GLOBAL VIEW - RESNET MODEL NOTEBOOK.ipynb

22 - Additional Model Information.md

23 - MiniRocket Model Notebook.ipynb

24 - SECTION 5 - readme.md

25 - SECTION 6 - readme.md

27 - SECTION 7 - readme.md

28 - Full Global View File.csv

29 - Full Raw View File.csv