Reconstructing the Public CALNet Baseline for Eclipsing Binary Classification. A Deep-Learning Model for Identifying Eclipsing Binaries

Research conducted by James D. Baker • mentored by Dr. Iqram Hussain

Abstract

CALNet is a deep-learning pipeline built to identify eclipsing binaries in TESS light-curve data. Eclipsing binaries are pairs of stars whose brightness changes in a repeating pattern when one star passes in front of the other. Establishing a reliable baseline for this task is important because future classifier extensions need a stable starting point. This project asks how much of the public CALNet baseline can be reconstructed, implemented, and run faithfully enough to support later model development.

To answer that question, I used the published paper, public repository, and author clarification to reconstruct the positive-sample manifest, generate a conservative negative candidate pool, stage a TESS sectors 1–10 slice, preprocess the data, and rebuild CALNet-style light-curve and GLS training inputs. I then ran the unchanged public baseline path on Google Colab Pro GPU, changing only the compute environment so the training could finish more reliably.

The completed 30-epoch run on 11,553 reconstructed rows achieved 95.18% accuracy, 94.72% recall, 95.63% precision, and 95.18% F1. These results show that a meaningful public-baseline reconstruction of CALNet is possible and can produce strong classification performance on a reconstructed slice. At the same time, exact historical reproduction of the final published model is still limited by missing artifacts, including the original negative-manifest workflow, original checkpoint, and full iterative retraining history. This reconstruction gives us a credible baseline for future scaling experiments and later controlled classifier extensions.