StarEmbed: Benchmarking Time Series Foundation Models on Astronomical Observations of Variable Starts

The first benchmark to test the state-of-the-art TSFMs on stellar time series observations ("light curves").

A complete benchmark framework for astronomical time series. This repository includes tools for (1) preprocessing raw light curves, (2) generating embeddings (with TSFMs and Astromer), (3) engineering handcrafted features, and (4) comprehensive evaluations on clustering, classification, and out-of-distribution detection.

| 🏠Benchmark Page | 🤗Huggingface Dataset | 📖Paper |

Directory Overview

src/datasets/

Raw light curve preprocessing and data preparation scripts
→ See datasets/README.md for detailed preprocessing workflows

src/model/

Time series foundation model implementations and embedding generation

• Astromer 1&2: Transformer-based astronomical time series model
• Chronos: Amazon's forecasting foundation model
• Moirai: Salesforce's universal time series model
• compute_avg_embeddings.py: Generate combined embeddings from multi-band data

src/benchmark/

Evaluation pipeline with pre-computed embeddings

• Classification: kNN, Linear models, MLPs, Random Forest with HPO
• Clustering: K-Means, hierarchical clustering, t-SNE visualization
  → See benchmark/README.md for complete evaluation workflows

bash_script/

job scripts for evaluation with hyperparameter search and multi-run script

---

Quick Start

1. Preprocess data: datasets/ → Raw light curves to standardized format
2. Generate embeddings: model/ → Extract features using TSFMs
3. Create combined embeddings: model/compute_avg_embeddings.py → Multi-band aggregation
4. Run evaluations: benchmark/ → Classification, clustering, visualization

---

License

All the code are under MIT license.