ATAT

Part of the light-curve family of open-source tools for astronomical time-series analysis.

Available from Python via the light-curve package: pip install light-curve. Documentation: light-curve.snad.space.

Paper

Becker, I., Pignata, G., Förster, F., Estévez, P. A., Cabrera-Vives, G., Vera, E., Carrasco-Davis, R., Astorga, N., Sanchez-Saez, P., Catelan, M., Cortés, C. C., de Jaeger, T., Pezoa, F., & Reyes, I. (2024). ATAT: Astronomical Transformer for time series And Tabular data. Astronomy & Astrophysics, 691, A163.

@article{atat2024,
  author  = {Becker, I. and Pignata, G. and F{\"{o}}rster, F. and
             Est{\'e}vez, P. A. and Cabrera-Vives, G. and Vera, E. and
             Carrasco-Davis, R. and Astorga, N. and Sanchez-Saez, P. and
             Catelan, M. and Cort{\'e}s, C. C. and {de Jaeger}, T. and
             Pezoa, F. and Reyes, I.},
  title   = {{ATAT}: Astronomical Transformer for time series And Tabular data},
  journal = {Astronomy \& Astrophysics},
  year    = {2024},
  volume  = {691},
  pages   = {A163},
  doi     = {10.1051/0004-6361/202451418},
}

Original code

https://github.com/alercebroker/ATAT (git submodule at models/atat/code/)

License

Apache-2.0 (Copyright 2026 ALeRCE Collaboration) — matching the upstream alercebroker/ATAT repository, which adopted a license following alercebroker/ATAT#2. See LICENSE.

Model overview

ATAT is a Transformer-based encoder for irregularly-sampled, multi-band astronomical light curves. The light-curve branch processes all six photometric bands jointly: each band's observations are independently embedded via a learned time modulation (sinusoidal Fourier basis), then all bands are merged, sorted by observation time, and passed through a 3-layer multi-head self-attention transformer. A learnable CLS token is prepended; its output at position 0 is the default representation. ATAT was trained for transient classification on the ELAsTiCC simulation (20 classes, LSST-like photometry).

Default configuration: 3 attention layers, 4 heads, head dimension 48 (d_model = 192), up to 65 observations per band, 6 bands, embedding dimension 192.

Input data format

The model was trained on the ELAsTiCC simulation dataset which emulates LSST photometry in 6 bands (u, g, r, i, z, Y). Each light curve is represented as a multi-band time series of flux measurements.

Preprocessing steps

Prepare inputs using the same pipeline applied during training (get_lc_md.py):

1. Split by band. Separate observations into 6 per-band sequences in the order [u, g, r, i, z, Y]. Each band gets an independent sequence of (time, flux) pairs.

2. Pad or downsample to 65 per band. If a band has fewer than 65 observations, right-pad with zeros to length 65. If a band has more than 65 observations, downsample to 65 by selecting indices linspace(0, n−1, 65) (uniform subsampling, not truncation). The result is data [65, 6] and time [65, 6]; padding positions hold 0.0.

3. Set the mask. Set mask = 1 for every slot containing a real observation and mask = 0 for every padding slot. This is a direct validity indicator — set it based on how many real observations each band has, not derived from flux values.

4. Shift time to start at zero. Subtract the minimum observed time (across all bands, ignoring padding slots) from all valid time entries. Padding time slots remain 0.0. Supply time in days; the model's internal time modulator uses T_max = 1500 days, so it is calibrated for light curves spanning up to roughly four years.

5. No flux normalisation. Pass raw flux values without any normalisation. The model was trained on SNANA FLUXCAL with reference zero point ZP = 27.5 (a source at 27.5 AB mag has FLUXCAL = 1). Inputs from a different photometric system or flux scale are outside the training distribution and may produce poor embeddings.

Inputs (ONNX)

Tensor	Shape	Description
data	[batch, 65, 6]	Per-band flux, SNANA FLUXCAL (ZP = 27.5), no normalisation
time	[batch, 65, 6]	Per-band observation times in days, shifted so earliest valid observation = 0; padding slots = 0
mask	[batch, 65, 6]	1 = valid observation, 0 = padding

Outputs (ONNX)

Single file atat.onnx with three named outputs:

Output name	Shape	Description
token	[batch, 192]	CLS token at position 0 after transformer (used in the paper)
mean	[batch, 192]	Masked mean pooling over per-observation features
sequence	[batch, 65×6, 192]	Per-observation features (CLS token excluded)

Request only the output(s) you need via session.run(["token"], feed) — onnxruntime will prune unused computation.

ONNX opset: 13.

Weights

Source: Google Drive — results_paper.zip Training dataset: ELAsTiCC (DESC LSST simulation, 20 transient classes) Checkpoint: results/lc/Exp_cfg_-arch=lc-seed=0/