PS7 - Data, Datasets, and Extraction Methods Context

This document consolidates the data, datasets, and extraction methods context for Problem Statement 7 (AI-enabled detection of exoplanets from noisy astronomical light curves) for the Bharatiya Antariksh Hackathon 2026.

1. Problem Core and Key Concepts

A planet crossing its host star blocks a fraction of the star's light. The main challenge is to extract this weak, periodic dip from noisy light curves, and then distinguish true transiting planets from false positives (impostors like eclipsing binaries, stellar spots, or detector artefacts).

Physics of the Transit

A transit is characterised by three geometric parameters:

  1. Depth ($\delta$): The fractional drop in brightness. It is proportional to the ratio of the projected areas:

$$\delta = \frac{\Delta F}{F} = \left(\frac{R_p}{R_\star}\right)^{2}$$ where $R_p$ is the planet radius and $R_\star$ is the star radius. For a Jupiter-size planet, $\delta \approx 1\%$. For an Earth-size planet, $\delta \approx 0.0084\%$ (84 ppm), which is often below the per-point noise floor.

  1. Duration ($T_{14}$): The time from first to last contact, determined by orbital speed and the impact parameter $b$ (how centrally the planet crosses the star):

$$T_{14} \approx \frac{P}{\pi} \frac{R_\star}{a} \sqrt{1-b^{2}}$$

  1. Period ($P$): The orbital period (the time between repeating dips), which determines the semi-major axis $a$ via Kepler's third law.

The Four Target Classes

2. Dataset Information

The data environment for PS7 involves several components, ranging from massive public catalogs to curated training sets:

Raw and Sector Light Curve Data

TESS Input Catalog (TIC)

Curated Labelled Dataset for Training

3. Extraction Methods and Pipeline Architecture

The recommended pipeline splits the problem into two distinct layers: Layer 1 for signal detection and extraction, and Layer 2 for machine learning classification.

+------------------+     +------------------------+     +--------------------------+
|  Raw Light Curve | --> |   Layer 1: Detection   | --> | Layer 2: Classification  |
|  (Time Series)   |     | (Detrend & Period Search)|     | (Fold, Feature Eng & ML) |
+------------------+     +------------------------+     +--------------------------+

Layer 1: Detection and Preprocessing (Signal Processing)

  1. Detrending (Flattening):
  1. Periodicity Search:
  1. Phase-Folding:

Layer 2: Feature Engineering and Classification (Machine Learning)

  1. Feature Extraction: Instead of training a deep neural network directly on raw folded light curves (which often fails to generalise from simulations to real noisy data), the recommended approach is a feature-based classifier. The most discriminating features extracted from the folded curve and stellar metadata are:
  1. Classifier Model:

Post-Classification Validation

  1. Parameter Fitting (Forward Modeling):
  1. Statistical Validation:
  1. Injection-Recovery Grid: