XFinder

Here is a rewrite of your project write-up that sounds like an enthusiastic 9th grader wrote it for a school science fair or a hackathon:

Inspiration Finding exoplanets—which are basically planets outside our solar system—is one of the coolest parts of astronomy right now. Space telescopes like NASA’s Kepler have collected so much data that scientists literally can't look through it all by hand anymore. We wanted to build an AI tool that could quickly scan through all this space data, spot potential new planets, and tell us which ones are worth checking out. Using machine learning makes this whole process way faster so astronomers can find new worlds without going crazy staring at spreadsheets!

What it does Our project, XFinder, works like an intelligent assistant for astronomers. It looks at space objects caught by the Kepler telescope and sorts them into three categories: Confirmed Exoplanet (an actual planet!), Candidate Exoplanet (looks promising, needs a closer look), or False Positive (a total false alarm). By weeding out the fake signals automatically, XFinder saves scientists tons of time.

How we built it We built a step-by-step machine learning pipeline to get the job done:

Checking Out the Data: We downloaded Kepler’s dataset from NASA and did some initial exploring to see what features we had, what numbers were missing, and how many planets versus false alarms were in the mix.

Cleaning Things Up: Raw space data is pretty messy! We filled in missing values, scaled the numbers so the big ones wouldn't mess up our math, and used a trick called SMOTE to fix a huge problem: there were WAY more false alarms than real planets in our dataset.

Splitting the Data: We saved 70% of our data to train our AI, 15% to tweak it, and held back 15% to test if it actually worked on stuff it hadn't seen before.

Training 7 Different Models: We tested a bunch of different algorithms to see which worked best—ranging from standard decision trees and logistic regression to advanced models like XGBoost and a 1D Convolutional Neural Network (a deep learning model usually used for patterns).

Combining Powers (Ensemble): We created a "Voting Classifier" that takes the opinions of all 7 models and combines them into one final, super-smart vote.

Scoring the Results: We checked how well each model did using metrics like accuracy and F1-score (which checks if the model is good at catching actual planets without throwing out too many fake guesses).

Challenges we ran into Missing Numbers: A lot of rows were missing info, especially error measurements. Deciding whether to delete those columns or guess the missing numbers took a lot of trial and error.

Way Too Many False Alarms: Because fake planet signals are way more common than real planets, our AI kept wanting to just guess "False Positive" every time to be safe. Setting up SMOTE to generate synthetic examples of real planets fixed this.

Hard Math vs. Clear Explanations: Super advanced models like XGBoost got great test scores, but it's hard to explain why they picked what they picked. Simpler models were easier to understand, so we had to balance accuracy with interpretability.

Accomplishments that we're proud of 7 Models in One Project: We successfully coded and compared seven totally different ML models, including a deep neural network!

Fixed the Imbalance Problem: Using SMOTE actually worked! Our models got way better at spotting the rare planet candidates instead of ignoring them.

Proving Real Science Features Matter: When we looked at what features the AI cared about most, it highlighted transit duration, transit depth, and orbit length—which are the exact same physical traits human astronomers look for!

It Works reliably: Our final Voting Classifier and XGBoost models scored super high accuracy scores across the board.

What we learned Data Cleaning > Fancy Models: We realized that spending time cleaning up missing values and fixing class imbalances makes a much bigger difference than just picking a fancy AI model.

Teamwork Works for AI Too: Grouping models together into an ensemble usually beats relying on just one single model.

Deep Learning isn't Just for Pictures: Neural networks can actually do a great job on tabular data (like spreadsheets) if you structure the inputs right!

Understanding the Science Helps: Knowing what the data actually meant in real life helped us figure out if our models were giving sensible answers or just making lucky guesses.

What's next for XFinder Fine-Tuning Hyperparameters: We want to use automated tuning tools (like Optuna) to squeeze every last bit of performance out of XGBoost and our neural network.

Creating New Features: We want to create custom features—like comparing a planet’s size directly to its star's size (planet_radius / stellar_radius)—to give the AI better physical clues.

Better Explanations with SHAP: We plan to add SHAP visual tools so when the AI flags a new candidate, it can show astronomers a chart explaining why it thinks it's a planet.

Building a Real Web App: Turn this project into a real-time web tool or API so astronomers around the world can upload Kepler data and test it instantly!