Self-Supervised Learning on Stellar Light Curves

First-author paper, ML4PS @ NeurIPS 2026. A self-supervised encoder pre-trained on unlabelled TESS light curves adds to engineered features on 7 of 10 tasks.

Role: First author, with Y. Ma and Prof. C. Theissen (UC San Diego)  ·  Stack: Python, PyTorch, scikit-learn, Weights & Biases  ·  Year: 2026 – Present

Paper: What does self-supervision add to engineered features for stellar light curves? — accepted at the Machine Learning and the Physical Sciences (ML4PS) workshop @ NeurIPS 2026 (arXiv:2610.14283).

arXiv → GitHub →

Question

The TESS space telescope produces far more stellar light curves than anyone can label by hand. Astronomers already have a few dozen hand-engineered features for them, and those features remain a strong baseline. Earlier work trained encoders on light curves and judged them on their own scores, but did not ask what they add to the features astronomers already use. That is the question this project answers.

Method

  • Self-supervised pre-training: a convolutional encoder trained on unlabelled TESS light curves to predict each window’s latent representation from its neighbours (GRU latent prediction, PyTorch).
  • Frozen evaluation: the encoder is frozen; its embedding is concatenated with 25 engineered features and read out with one fixed linear probe on 10 downstream tasks (variability classes, oscillation targets, rotation period).
  • Controls: the same probe on the engineered features alone, on the embedding alone, on an untrained encoder, and two networks trained end-to-end on the same labelled stars (a Conv1D and an MLP).
  • Statistics: multi-seed runs with bootstrap confidence intervals, tracked in Weights & Biases.

Result

  • Adding the learned features improves the linear probe on 7 of 10 tasks over the engineered features alone; an untrained encoder does not.
  • The combination also beats both end-to-end networks on most tasks.
  • Takeaway: an encoder trained without labels gives astronomers features that complement the ones they already use, rather than replace them.
Paper Figure 1. (a) Gain from adding the learned features to the engineered features, linear readout (solid) and XGBoost readout (faded). (b) The fused probe against two networks trained end-to-end. Bars are 2 SE.