Hey Dev Community! I am so excited to share a major pivot in my research journey. While my previous work (2025–2026) was all about the digital world—specifically detecting misinformation in high-variance network environments—I recently started asking myself a big question: Is the extraction of truth from noise a universal mathematical challenge?
To find out, I looked away from the screen and up at the stars! 🌌
I’ve just released my latest project, AstroNet-Lite, a dual-path Convolutional Neural Network designed to find exoplanets in the chaotic, high-noise data from NASA’s TESS satellite.
What Makes This Research Different?
If you've followed my "Edge NLP" work, you know I’m obsessed with lightweight, hardware-aware optimization. This time, I took those same principles and applied them to Astrophysics. Here is how this phase of my work stands out:
A New Kind of "Noise": Instead of network variance, I’m now battling photon noise, stellar variability, and instrument-induced trends that mask the tiny 1% dip in light caused by an exoplanet.
The Dual-Scale Architecture: Unlike monolithic classifiers, AstroNet-Lite uses two distinct convolutional paths to decouple spatial features:
The Global View: Uses large kernels (size 7) to understand the "Star" and its natural cycles.
The Local View: Uses small kernels (size 3) to "Zoom" in on the "Planet," identifying the sharp entry (ingress) and exit (egress) points of a transit.
Extreme Efficiency: I’m proving that "Big Science" doesn't need "Big Compute". This model is only 312 KB—smaller than a single high-resolution photograph!
Real-World Validation: I successfully navigated the "Sim-to-Real" gap. After training on synthetic physics-based data, I used Savitzky-Golay detrending to identify actual confirmed planets in NASA’s archives, like TOI-700 d and WASP-126 b.
Why This Matters 🌍
Operating within the bandwidth constraints of rural Tripura, I wanted to show that the search for another Earth isn't just for government agencies with supercomputers. By democratizing these tools, any student with a curious mind and an efficient algorithm can join the frontier of discovery.
Whether it's a deceptive claim in a digital feed or a planet orbiting a distant sun, the engineering challenge is the same: distinguishing the signal from the noise.
Special Thanks & Resources
A huge thank you to the open-science community and the NASA TESS archive for making this data accessible to independent researchers everywhere. This journey from NLP to Astrophysics has been a whirlwind, and I'm so grateful for the support!
Read the New Paper:
AstroNet-Lite: A Dual-Scale Convolutional Framework for Automated Exoplanet Discovery (https://doi.org/10.5281/zenodo.18405183)
Check Out My Previous Works:
Democratizing Truth: Optimizing Transformer Models for Client-Side Misinformation Detection (https://zenodo.org/records/17879430)
Neural Network Quantization for Edge Deployment — Field Validation (https://doi.org/10.5281/zenodo.18140944)
I'd love to hear your thoughts!
Top comments (1)
I'm not sure, but it seems to me that this is just the kind of project that could unlock so many new discoveries about our universe. I'm fascinated by the idea that someone's "Big Compute" setup can help us discover new worlds. It's mind-boggling to think that with this AstroNet-Lite model, we might stumble upon entirely new planets just by processing the vast amounts of data from TESS. And the fact that it's so lightweight - 312 KB, wow! That just feels like a testament to human ingenuity.