Prediction of terrestrial radiation belts using deep learning

The Earth’s Radiation Belts and their risks

The terrestrial Van Allen radiation belts constitute a region of high-energy particles trapped in the Earth’s magnetic field. These particles arrive towards the Earth following a sudden solar eruption or as background radiation from the outer space. 

The radiation belts consist of an inner zone, mainly populated by protons, and an outer region, mostly filled by electrons, even if both particle types coexist in both zones. These two belts are separated by a slot layer. However, this whole configuration is very dynamic, especially during high solar activity.

The continuous monitoring of near-Earth radiation is critically important. Sudden increases of particle fluxes and prolonged radiation exposure are serious threats to both human infrastructures and health. Essential systems at risk include power grids, communication satellites and aviation equipment, which can experience disruptions or complete failure. 

Moreover, humans exposed to such radiation face risks ranging from acute injuries to long-term health effects, including cancer.

The prediction model

A deep learning Neural Network Long Short-Term Model (LSTM) has recently been developed in a prototyping stage that provides forecast of the Van Allen radiation belts dynamics [Botek et al., 2023]. The model is driven by the Energetic Particle Telescope (EPT) instrument satellite data [Pierrard et al., 2014]. 

This instrument was launched onboard the European Space Agency PROBA-V satellite in May 2013 on a Low Earth Orbit at 820 km. 

The EPT discriminates between electrons, protons and helium ions in energy ranges 0.5-20 MeV, 9.5-300 MeV and 38-1200 MeV, respectively. The high spatial and time resolution of the measurements during more than 11 years represents an extraordinary coverage of quiet and stormy solar activity periods. 

These long-term observations allow the successful training of a robust prediction model capable of very good performances from 1 to 8 L-shells ranges, where L is the equatorial distance of the magnetic field lines, expressed in Earth radii (R_E). 

The results of the model are the predictions of the logarithm of electron fluxes for two energy channels: 550 keV (labeled as ‘e1’) and 1.7 MeV (labeled as ‘e5’).

Figure 1 displays the performance of the ‘e5’ electron fluxes predictions during a period in 2018 along the PROBA-V orbit. The three panels represent different L-shell ranges: a) containing all the radiation belts regions (1-8 R_E), b) removing the inner belt (2-8 R_E) and c) only considering the outer belt (3-8 R_E). The model uses as inputs: the satellite observed fluxes, the satellite coordinates and SYM/H geomagnetic index (measure of the Earth’s surface magnetic disturbances). These are referred as ‘input group 4’. 

Figure 2 illustrates the comparison of the 1-hour resolution model predicting the ‘e1’ electron flux evolution all over the L-time spectrum considering only the outer radiation belt layer. All the other settings are as in Figure 1.

Ongoing further model developments encompass:

1. predictions at higher spaciotemporal resolutions to capture phenomena at different scales and 
2. physics informed integration into the learning procedure.

References:

• Botek, E., Pierrard, V., Winant, A. (2023). Space Weather, 21, e2023SW003466. https://doi. org/10.1029/2023SW003466.
   
• Pierrard, V. et al. (2014), The Energetic Particle Telescope: First results, Space Science Rev., 184(1), 87-106, DOI: 10.1007/s11214-014-0097-8, 2014.