Hi everyone,
I introduced myself briefly when I first joined the AWG, but to reintroduce myself: my name is Udayan. I am an undergraduate Computer Engineering student and ML research enthusiast based in Pune. I attended the recent OSDR AI/ML meeting and was advised to bring this question to the forum for a deeper discussion.
I am currently developing an automated exoplanet detection pipeline using TESS high-cadence Target Pixel Files (TPFs). The core issue I am tackling is the high rate of false positives caused by stellar blending and eclipsing binaries in crowded fields.
To address this, rather than relying on a pure black-box classifier, I am building a “Detection & Vetting” architecture:
- The AI Scanner: A Bayesian 3D-CNN that processes the spatiotemporal TPF data to flag periodic dips, using Monte Carlo Dropout to generate uncertainty scores.
- The Physics Validator: A verification loop, constrained by a Physics-Informed Neural Network (PINN), that forces every AI-flagged candidate to pass strict physical checks like Centroid Shift Analysis and Odd-Even Depth tests.
For those of you who have experience building and testing models on highly noisy, complex NASA datasets, what is your best advice for rigorously validating a hybrid AI-physics pipeline like this?
I plan to benchmark it against the confirmed TESS Objects of Interest (TOI) catalog, but I would greatly appreciate any insights on best practices for proving out the physics-validation layer, or pitfalls to avoid when handling extreme instrumental artifacts before the data hits the AI.
Thanks for your time and any advice you can share!
Udayan