LAUREN

Real-time auroral electrojet indices. LAUREN predicts the AU and AL indices continuously, on the scale of the classic observatory-based indices. It is trained on ground magnetometer observations from the INTERMAGNET network and on GeoDGP ground magnetic perturbation predictions driven by the MIDL solar-wind merge, together with the Dst index and F10.7 solar flux. All of these inputs are fed to the model in real time, and it runs continuously with a short-range forecast ahead of real time.

In development

↔ Drag to pan

Loading…

Red lines are the LAUREN prediction — AU above zero (the eastward auroral electrojet), AL below (the westward); the dotted continuation right of the marker is the forecast, revised continuously as new observations arrive. Black dashed lines are the observed indices where published (WDC Kyoto quicklook, roughly two weeks behind real time). Click and drag to pan back through the last 35 days at full one-minute resolution; double-click snaps back to the selected window. The plot refreshes every five minutes.

Validation — daily correlation r

Validation — daily RMSE (nT)

How LAUREN works

LAUREN is a machine-learning model that estimates the AU and AL indices for every minute. AU and AL measure the strength of the eastward and westward auroral electrojets, the electric currents that flow in the ionosphere along the auroral oval. The model uses LightGBM, a method that adds up the outputs of many decision trees. There are two models, one for AU and one for AL, each with 600 trees. For each minute they take these inputs:

GeoDGP’s real-time prediction of the north component of the ground magnetic perturbation at 5,580 points, every 2° in latitude and longitude from 30°N to the pole. GeoDGP is a separate model that predicts ground magnetic perturbations from the solar wind.
The solar wind from MIDL’s real-time merge, moved forward in time from L1 to the bow-shock nose: the three components of the magnetic field (GSM), the flow speed, the proton density, and the temperature. Each enters as its current value and as twelve 5-minute medians covering the past hour.
The Dst index from WDC Kyoto and its hour-to-hour changes over the previous 12 hours.
The F10.7 solar radio flux.
Ground magnetometer data from 30 INTERMAGNET observatories: the horizontal (H) component minus a quiet-day baseline, as six 5-minute medians reaching back 25 minutes from the latest data; each station’s magnetic local time; and how many minutes old its latest data are. If a station is not reporting, the model runs without it.
The tilt of Earth’s magnetic dipole toward or away from the Sun, the time of day (UT), and the day of the year.

The models were trained on the 18 years from 2005 to 2022 to match the observed AU and AL indices from the World Data Center for Geomagnetism, Kyoto (1-minute values from NASA OMNI). For those years GeoDGP was rerun with MIDL solar wind, so the training inputs are made the same way as the real-time ones. The magnetometer data were given delays drawn from the delays measured on the real-time feed. Minutes with strong AL counted more in training (weight 1 + |AL|/300 nT, at most 8), so that quiet times do not outweigh storms. The models were then tested on 2023 and 2024, which they never saw in training, including the May 2024 storm. The correlation with the observed indices was 0.884 for AU and 0.888 for AL, and the root-mean-square error was 38 nT for AU and 71 nT for AL.

Solar wind measured at L1 takes roughly 30 to 60 minutes to reach Earth. Because the solar-wind and GeoDGP inputs are available before that wind arrives, LAUREN gives AU and AL up to about 45 minutes ahead of the current time, less when the solar wind is fast.

← All products About LAUREN →