What does it do or how does it work?

ExoSeeker is a platform for creating custom machine learning models for predicting whether a Kepler Object of Interest could be a potential exoplanet. A user can upload new data, freely build their own model by selecting from multiple estimators (random forest, gradient boosting, and multi-layer perceptron) and adjust each one's primary hyperparameters. Once the user has decided to train their model, the uploaded dataset will be preprocessed automatically. Dropping unnecessary columns and ensuring to keep essential ones, such as the KOIs' transit properties. Missing values would be filled with their features' means. The preprocessed data is split for training and testing and then fitted in the custom model, where the chosen estimators are stacked and their output is passed as input to a final estimator. The model would predict the exoplanet disposition on the test dataset and compare predicted classifications to the actual classifications.

ExoSeeker Interface

The built model's performance is measured in: accuracy, the percentage of data predicted correctly; sensitivity, the percentage of data predicted correctly that belongs to positive (correctly predicted) classes; specificity, percentage of data predicted correctly that belongs to negative (wrongly predicted) classes; precision, percentage of correct positive predictions made; and the F1 score, a performance metric derived from the precision and recall of the model. A confusion matrix is also generated with the following values: true positives, exoplanets confirmed and classified as confirmed; true negatives, exoplanet candidates and classified as candidates; false positive, exoplanet candidates but classified as confirmed; and false negatives, exoplanets confirmed but classified as candidates.

Model Evaluation

These values were obtained by comparing the predicted exoplanet disposition to the actual values in the test dataset. The model is then saved in the local machine, and the user can upload a new dataset without the exoplanet disposition, with only the feature to run predictions on it using the saved model.

Sample Target Dataset (Opened in VSCode)

Generated Predictions (Opened in VSCode)

What benefits does it have?

It has a user-friendly web interface that allows anyone to create their own AI models for identifying new exoplanets.
Each component has a tooltip that provides useful information, such as instructions for uploading new data, properties of each estimator's main hyperparameters, and the description of each metric in the model evaluation.
ExoSeeker allows for stacking multiple estimators, enabling the harnessing of the strength of one to increase the model's performance.
The user can mix and match multiple models, tweak each one's main hyperparameters, empowering the user to create fully customized machine learning models.

What is the intended impact of the project?

The goal of ExoSeeker is to empower anyone, whether they're a researcher or an enthusiast, to easily and swiftly create sophisticated AI models for discovering new exoplanets. By streamlining data processing of objects of interest, the testing of unique machine learning models, evaluating them, and acquiring predictions for unseen new data, this project hopes to contribute to the ongoing research of finding new worlds.

What tools, coding languages, hardware, or software did you use to develop your project?

Tools: VSCode, Jupyter Notebook
Hardware: CPU 12th Gen Intel(R) Core(TM) i7-12650HX
Coding Language: Python
Software: numpy, pandas, sklearn, altair, matplotlib, seaborn, streamlit

How is your project creative?

ExoSeeker has a unique approach of utilizing multiple machine learning models to generate predictions and building them without writing code.

What factors did your team consider?

The main factors that were taken into account were usability and accessibility of the user interface, the level of customization of the machine learning model, and the target data for identifying exoplanets.

2025 NASA Space Apps Challenge

ExoSeeker