Rare Transit Pipeline

Rare transits are needle‑in‑a‑haystack events: asymmetric dips, ring‑like overshoots, or sequences of shallow events buried in stellar variability and instrument systematics. Manual vetting doesn’t scale, so we built an end‑to‑end system that makes these signals pop without sacrificing statistical validity.

1) Data prep that respects the physics

We pulled Kepler/K2/TESS light curves (Lightkurve/Astroquery), removed long‑term trends with Gaussian‑process fits, and windowed residuals into ~36–60h segments. We handled gaps and reaction‑wheel artifacts, normalized windows, and split by star so evaluation never leaks across targets.

2) A label‑free encoder that understands morphology

A compact self‑supervised model learned to reconstruct masked residuals while matching a whitened power spectrum. The result is a mission‑agnostic embedding that highlights ingress/egress‑like structure without needing hand labels.

3) Physics‑informed detection

We run fast, PSD‑aware matched filters using physically motivated templates (exocomet tails, ringed trapezoids, disintegrator combs, circumbinary sequences). Each window gets a set of GLRT‑style scores plus template parameters.

4) Calibrated scoring you can trust

Scores from the encoder and detectors are mapped through isotonic regressions and converted to conformal p‑values using a large null pool. We then control false discovery rate (FDR) at star and global levels, so a “significant” call means what it should.

5) Evidence, not just numbers

For every candidate we generate an image card that overlays the detrended flux, matched‑filter responses, residual checks, and metadata. The explorer you can open below renders these precomputed artifacts with zero backend code.