griffingoodwin04/FOXES-Data

FOXES Dataset: SDO/AIA EUV Images + GOES Soft X-ray Flux

Pre-processed, pre-split multiwavelength EUV image stacks paired with simultaneous GOES soft X-ray (SXR) irradiance measurements. This dataset was created to train and evaluate FOXES (Framework for Operational X-ray Emission Synthesis), a Vision Transformer-based model that predicts solar SXR flux from spatially-resolved EUV imagery.

Code & model: griffingoodwin04/FOXES

Model card: griffingoodwin04/FOXES-model

---

Dataset Summary

Each sample is a single timestamp pairing:

• AIA image stack — 7-channel EUV image from SDO/AIA, one channel per wavelength band, normalized to [-1, 1]
• SXR flux value — simultaneous GOES B-channel (1–8 Å) soft X-ray irradiance in W/m²

The dataset covers multiple years of solar observations and is split by date into train, validation, and test sets to prevent temporal leakage.

---

Dataset Structure

Features

Column	Type	Shape	Description
filename	string	—	ISO 8601 timestamp used as a unique sample ID (e.g. 2012-03-07T00:12:00)
aia_stack	float32	(7, 512, 512)	Multi-wavelength EUV image stack, pixel values in [-1, 1]
sxr_value	float32	scalar	GOES XRSB flux in W/m² (raw physical units, not log-normalized)

AIA Wavelength Channels

The 7 channels in aia_stack correspond to SDO/AIA EUV bandpasses in the following order:

Index	Wavelength (Å)
0	94
1	131
2	171
3	193
4	211
5	304
6	335

Splits

Split	Local directory
train	train/
validation	val/
test	test/

---

How to Use

With the FOXES pipeline (recommended)

The FOXES repo provides a downloader that reconstructs the expected local directory layout:

git clone https://github.com/griffin-goodwin/FOXES.git
cd FOXES
pip install -r requirements.txt

python download/hugging_face_data_download.py \
    --config download/hf_download_config.yaml

Configure output paths in download/hf_download_config.yaml:

repo_id: "griffingoodwin04/FOXES-Data"
aia_dir: "/your/data/AIA_processed"
sxr_dir: "/your/data/SXR_processed"
splits:
  - train
  - validation
  - test

This produces:

/your/data/
├── AIA_processed/
│   ├── train/       ← .npy files, shape (7, 512, 512) float32
│   ├── val/
│   └── test/
└── SXR_processed/
    ├── train/       ← .npy files, scalar float32
    ├── val/
    └── test/

With the HuggingFace datasets library

from datasets import load_dataset
import numpy as np

ds = load_dataset("griffingoodwin04/FOXES-Data", split="train", streaming=True)

for sample in ds:
    aia = np.array(sample["aia_stack"], dtype=np.float32)   # (7, 512, 512)
    sxr = float(sample["sxr_value"])                         # W/m²
    t   = sample["filename"]                                  # ISO timestamp string
    break

Downloading a subset

from datasets import load_dataset

# Stream and take only 1000 samples from validation
ds = load_dataset("griffingoodwin04/FOXES-Data", split="validation", streaming=True)
ds = ds.shuffle(seed=42, buffer_size=3000).take(1000)

---

Acknowledgements

This work is a research product of Heliolab (heliolab.ai), an initiative of the Frontier Development Lab (FDL.ai). FDL is a public–private partnership between NASA, Trillium Technologies, and commercial AI partners including Google Cloud and NVIDIA.

This material is based upon work supported by NASA under award number No. 80GSFC23CA040.

Team: Griffin Goodwin, Alison March, Jayant Biradar, Christopher Schirninger, Robert Jarolim, Angelos Vourlidas, Viacheslav Sadykov, Lorien Pratt