HabitaX

🌌 HabitaX: Physics-Informed AI for Exoplanet Verification & Bio-Potential Assessment

Live Web App: https://habitax.streamlit.app/

"Somewhere in the Kepler dataset lies a signal that looks identical to Earth, but 90% of machine learning algorithms will mistake stellar jitter for a living world—and 9% will cheat by reading the post-hoc answer key. HabitaX was built to do neither."

When evaluators and NASA scientists read through hundreds of submissions, most look identical: off-the-shelf classifiers trained on raw tabular numbers, claiming synthetic 99% scores due to unnoticed data leakage. HabitaX breaks this mold. It acts like an astrophysicist: systematically purging false positive flags, embedding fundamental laws of planetary physics directly into model features, and offering transparent SHAP-based reasoning alongside a Bio-Potential Index for atmospheric target prioritization.

🎯 Inspiration & Problem Statement

The NASA Kepler Space Telescope discovered thousands of Objects of Interest (KOIs) by measuring subtle dips in starlight as planets cross in front of their host stars. However, ground-based telescope follow-ups are extremely expensive and time-consuming. Most candidate signals turn out to be false positives caused by eclipsing binary star systems, stellar variability, or background noise. Existing machine learning entries frequently suffer from two critical flaws:

1) Data Leakage Disqualification: Training on disposition flags (such as koi_pdisposition, koi_score, and koi_fpflag_*) gives artificial 99% accuracy that immediately collapses on real unseen telescope data.

2) "Black Box" Blindness: Predicting "confirmed" or "false positive" without physical context or explainability leaves astronomers unable to trust the output.

🔬 What HabitaX Does

HabitaX is a physics-informed AI platform designed to automate exoplanet signal verification and astrobiological assessment: Leakage-Safe Binary Classification: Classifies real exoplanets vs. false positives using only true observational features, dropping unresolved CANDIDATE rows to prevent label contamination during training. Physics-Informed Feature Engineering: Formulates planetary metrics including Calculated Planet Radius ($R_p$), Earth Similarity Index ($ESI$), and Insolation Flux relationships directly inside the feature pipeline.

Explainable AI (XAI): Employs SHAP (SHapley Additive exPlanations) to break down every single prediction into exact feature contribution weights. Bio-Potential & Habitability Screening: Ranks verified planets into habitability tiers to assist astrobiologists in selecting candidates for atmospheric spectroscopy (e.g., JWST follow-ups). Deployed Interactive Web App: Serves predictions, dynamic gauges, SHAP waterfall plots, and downloadable mission reports via Streamlit.

🛠️ Technical Architecture & Pipeline

Data Cleaning & Leakage Prevention Filtered from the official NASA Kepler Cumulative Dataset (kepler_dataset.csv). Strict removal of leakage features (koi_score, koi_pdisposition, koi_fpflag_nt, koi_fpflag_ss, koi_fpflag_co, koi_fpflag_ec). Imputed missing values using K-Nearest Neighbors (KNN) Imputation to preserve multi-variable physical correlations.
Astrobiology & Physics Feature Engineering Calculated Planet Radius ($R_p$): $$R_p = R_* \cdot \sqrt{\text{Transit Depth}}$$ Multi-Parameter Earth Similarity Index ($ESI$): Integrates planetary radius, equilibrium temperature ($T_{eq}$), and stellar flux ($S_{eff}$) relative to Earth baseline standards. Biological Potential Index: Evaluates surface temperature stability and insolation suitability to assign a decision-support habitability tier (Habitable, Sub-Habitable, or Non-Habitable).
Model Optimization & Training Core Classifier: LightGBM Gradient Boosting Classifier tuned using Optuna hyperparameter optimization. Validation Strategy: 5-Fold Stratified Cross-Validation to prevent class imbalance skew and guard against overfitting.

💻 Live Web App Features (habitax.streamlit.app) The web application provides an intuitive interface for researchers and students: Planet Analyzer: Input orbital period, transit depth, stellar radius, equilibrium temperature, and stellar magnitude. AI Verification & Confidence Meter: Instantly returns planetary candidate verification probability. SHAP Feature Explainer: Visualizes which physical parameters drove the AI decision toward "Confirmed" or "False Positive". Mission Summary Report: Generates exportable summary cards for classroom use or research presentations.

💡 Challenges & Future Roadmap Cross-Mission Domain Shift: Space telescopes operate under different noise distributions. Inspired by cross-mission challenges between Kepler and TESS data, our next step is integrating domain adaptation techniques to generalize HabitaX across TESS and James Webb Space Telescope (JWST) datasets. Direct Archive Sync: Building an automated REST API to pull new KOI signals directly from the NASA Exoplanet Archive for real-time verification.