Discovering New Worlds with Explainable AI
AkashDrishti combines astrophysics, adaptive noise analysis, Transit Least Squares, and machine learning to assist astronomers in detecting exoplanets from noisy stellar light curves.
Real-time Discovery Statistics
Aggregated across simulated Kepler, K2, and TESS observation windows processed by AkashDrishti.
Data Ingest Console
Drop a FITS or CSV light curve, or pick a sample mission target to begin analysis.
Drag & drop a light curve
Accepts .fits / .csv up to 200 MB
CSV needs time + flux columns.
First confirmed planet in the habitable zone of a Sun-like star.
Earth-cousin candidate orbiting an older Sun-like star.
Seven terrestrial planets; three in the habitable zone. Also observed by TESS.
First Earth-sized habitable-zone planet found by TESS.
Active vetting target used as the default demo light curve.
Runs the real preprocessing + TLS pipeline on this target and jumps to live results below.
Aitoff projection · Equatorial coordinates
Sky distribution of active candidates across Kepler, K2, TESS & TOI footprints. Hover a point for ID and coordinates.
Live Star Analysis
Connected to the AkashDrishti backend — runs real preprocessing + TLS on the requested star.
No analysis running yet — go to Data Ingest Console and enter a Star ID.
- 1Loading raw light curve
- 2Removing NaNs and outliers
- 3Normalizing flux
- 4Detrending (Wotan / Savitzky–Golay)
- 5Masking poor cadences
- 6Running TLS
- 7Running BLS
- 8Engineering features
- Coming soon — trained CNN/GBM model coming soon — heuristic baseline active9Running ML inference
- 10Applying confidence calibration
- 11Estimating MC-Dropout uncertainty
- 12Generating XAI explanations
- 13Composing candidate ranking
Explainable AI
Every classification is decomposed into physically meaningful evidence the astronomer can audit.
Regular Transit Periodicity
Hover to revealThree evenly spaced dips at P = 4.219 d with σ < 0.4%.
Strong Signal-to-Noise
Hover to revealSNR of 14.2σ exceeds the 7σ candidate threshold.
Stable Transit Depth
Hover to revealDepth = 1180 ± 32 ppm across all observed transits.
No Secondary Eclipse
Hover to revealPhase 0.5 search shows null result (Δ < 18 ppm).
Low Stellar Variability
Hover to revealDetrended RMS = 92 ppm, consistent quiet star.
Stable Centroid Position
Hover to revealCentroid drift < 0.02 px — no contaminating source.
High Detection Confidence
Hover to revealHybrid classifier returns 0.974 (calibrated).
Astronomer Recommendation
A single, actionable verdict synthesising all upstream evidence.
Likely Exoplanet · Confidence 97.4%
Three consistent transit events detected, with a transit shape matching a planetary signature and a low false-alarm probability. Centroid analysis rules out background eclipsing binaries.
End-to-End Architecture
A single, continuous pipeline from raw observation to vetted scientific report.
What Makes AkashDrishti Different
Hover any capability to flip the card and read how it's engineered into the pipeline.
Adaptive Noise-Aware Preprocessing
Hover to revealEngineered into the pipeline: classifies Gaussian, red, and instrumental noise per light curve, then routes it through the optimal denoiser — wavelet, Savitzky-Golay, or GP detrending.
Physics-Informed AI
Hover to revealEngineered into the model: Mandel-Agol transit priors, Kepler's third law, and limb-darkening constraints regularize the network so predictions stay astrophysically plausible.
Explainable Predictions
Hover to revealEngineered into every verdict: SHAP attributions and saliency maps surface exactly which transit features drove the score — no black-box dispositions.
Confidence Calibration
Hover to revealEngineered into the output layer: temperature scaling and isotonic regression ensure a stated 90% confidence really means 90% — critical for telescope-time triage.
Candidate Prioritization
Hover to revealEngineered into the ranking engine: multi-objective scoring weighs SNR, novelty, follow-up cost, and habitability to surface the highest-yield targets first.
Astronomer Decision Support
Hover to revealEngineered into the UI: side-by-side periodograms, phase folds, and centroid checks give domain experts everything they need to confirm or reject a candidate.
Human-in-the-Loop Design
Hover to revealEngineered into the feedback path: every Confirm / Reject / Escalate decision streams back into retraining, so the model improves with every astronomer interaction.
Scientific Report Preview
Every analysis produces a shareable, peer-reviewable PDF dossier.
Includes raw + detrended light curves, periodogram, phase-folded transit, residuals, vetting summary, SHAP explanation, and recommended follow-up program.
Team AkashVyuh
Harshit Kumar Ray
Defines mission scope, scientific objectives, and the validation framework that underpins every recommendation.
Harshit Kumar Ray
Karan
Owns the FastAPI pipeline end-to-end — light curve ingest, TLS/BLS detection, feature engineering, model inference, and deployment to production.
Karan
Rajnandini
Drives literature review, scientific writing, and the reproducibility of every reported result.
Rajnandini
Shreelekha
Builds the mission-control interface, plotting subsystem, and the report generation engine.
