Machine Learning Aids Jupiter Exploration: DiffFluxReconstruction Reconstructs Solar Energetic Particle Energy Spectra

The RADEM radiation monitor on the European Space Agency's JUICE mission uses an ensemble neural network model to reconstruct the differential energy spectra of solar energetic particles from count rate data via machine learning techniques, providing critical radiation environment monitoring capabilities for deep space exploration. This project (DiffFluxReconstruction) aims to solve the problem of energy spectrum reconstruction in Jupiter's extreme radiation environment and improve the accuracy of spacecraft radiation risk assessment.

In solar system exploration, the radiation environment is a core challenge for spacecraft design and mission planning. Solar Energetic Particle (SEP) events release high-energy charged particle streams (with energy ranging from several mega-electron volts to giga-electron volts), which pose threats to spacecraft equipment, solar panels, and astronaut health. The radiation environment of the Jupiter system is harsh; its magnetic field traps a large number of high-energy particles to form strong radiation belts. ESA's JUICE mission will delve into this environment to explore Jupiter's icy moons.

The RADEM radiation monitor aboard the JUICE mission is responsible for real-time radiation environment monitoring and measuring the count rate of incident particles. However, count rates cannot directly reflect particle energy distribution. Traditional energy spectrum reconstruction relies on simplified physical models and assumptions, which suffer from insufficient accuracy or low computational efficiency in the complex deep space environment.

DiffFluxReconstruction uses an ensemble neural network approach: the core consists of three neural network ensembles, each containing 20 independently trained networks. The ensemble strategy reduces prediction variance through multi-model voting and improves robustness. The three ensembles have different numbers of input features; during inference, they are dynamically selected based on input count rates and background thresholds to reduce noise impact and optimize accuracy.

Training data sources include RADEM's on-orbit data from October 21, 2023, to February 17, 2025, and radiation energy spectrum data from NASA's STEREO-A detector for cross-validation. The project includes response function modeling for each RADEM detection unit (a key link between count rates and physical energy spectra), and also develops a simulated dataset generation tool to expand samples or test performance under extreme conditions.

The project is developed using Python 3.14 with a modular code structure: Data (raw/preprocessed data), NeuralNetworks (model parameters/hyperparameters/simulated data), Response Functions (response function definitions), JupyterNotebooks (visual analysis), Modules (core functions), Scripts (training/tuning scripts), Tests (functional tests). A virtual environment configuration guide is provided to ensure reproducibility.

The project's result paper "Flux reconstruction with Machine Learning techniques for the ESA JUICE mission radiation monitor, RADEM" is under peer review at AGU journals. Its value includes: supporting JUICE's real-time radiation monitoring for risk assessment; demonstrating the potential of machine learning to outperform traditional methods in space data processing; and its open-source nature promotes community collaboration and can be referenced by other missions.

DiffFluxReconstruction is a typical case of combining machine learning with deep space exploration, providing important data support for the JUICE mission's deep exploration of the Jupiter system. As more deep space missions are implemented, such machine learning-assisted data analysis methods are expected to play a greater role in the field of space science.