surya-bench-flare-forecasting

What surya-bench-flare-forecasting is

The surya-bench-flare-forecasting dataset is a full-disk solar flare forecasting benchmark developed by the nasa-ibm-ai4science initiative and researchers at Georgia State University. It extracts and standardises labels from National Oceanic and Atmospheric Administration (NOAA) Geostationary Operational Environmental Satellite (GOES) flare events between May 2010 and December 2024. Solar flares are intense bursts of radiation resulting from magnetic reconnection in the solar atmosphere, classified into categories such as C, M, and X based on peak soft X-ray flux.

The benchmark addresses the challenge of standardising evaluations for heliophysics and space-weather machine learning models. By sampling full-disk observations at an hourly cadence across a 24-hour rolling forecasting window, it provides consistent target labels to evaluate predictions of forthcoming significant flaring activity.

How it was made

The dataset is curated from NOAA GOES soft X-ray flux event records covering 13 May 2010 to 31 December 2024. Observations are sampled every hour using a 24-hour rolling prediction window, yielding a total of 128,328 instances. For each 24-hour window, annotations record the maximum GOES flare class based on peak X-ray flux (or 'FQ' if quiet with no flares) and a cumulative flare index calculated as a weighted sum of subclasses for all flares of class C or greater, where C-class flares have a weight of 1, M-class a weight of 10, and X-class a weight of 100.

From these continuous measurements, two binary forecasting labels are derived: one flag marking whether the maximum flare class reaches or exceeds M1.0, and a second flag marking whether the cumulative flare index reaches or exceeds 10. The tabular data is partitioned into predetermined splits: training periods (15 February to 31 December for 2010–2019), validation periods (15–31 January for 2010–2019), testing periods (all instances from 2020–2024), and an additional leaky validation split containing the early-to-mid January and February intervals from 2010 to 2019.

Key facts

What you can use it for

Binary flare occurrence forecasting

Train and evaluate binary classification models to predict whether a flare of class M1.0 or greater will occur within the next 24 hours.

Cumulative flaring index regression

Develop regression pipelines to estimate total cumulative flare energy and frequency index across hourly rolling windows.

Ordinal flare-class classification

Assess multiclass or ordinal machine learning models tasked with predicting specific peak GOES flare classes (such as quiet, C, M, or X).

Explore the data

The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.

Strengths and limits

Questions