Zurück

Seminar: The application of machine learning methods in space physics

Alexander Lukin will discuss how machine learning and explainable AI can be used to analyze space physics data, predict plasma-sheet electron fluxes, and uncover the physical processes driving these predictions

Donnerstag 19.11.2026 01:11 Uhr

Machine learning has developed rapidly over the past decades, and many of its methods can be applied directly to space physics problems. Long-term measurements from solar wind monitors, missions such as THEMIS, Van Allen Probes, MMS and Arase, and ground-based observatories provide large datasets for training complex models. These data and methods are now used for a broad range of tasks, including time-series forecasting of space weather, building global empirical models, detection of events such as interplanetary coronal mass ejections, classification of solar and auroral images, and as fast surrogates for computationally expensive simulations. With the development of GPU computing, classical methods such as linear regression, SVM and gradient boosting have been joined by deep learning: alongside fully connected neural networks, convolutional, recurrent and transformer architectures are now widely used. Many pretrained models have appeared in recent years that can be adapted to specific space physics problems with little or no fine-tuning. Also, for a long time machine learning models were treated as black boxes: their predictions were hard to interpret in terms of the input parameters, and the influence of each driver remained unclear. Explainable AI (XAI) is changing this. Methods such as SHAP values, integrated gradients and permutation importance trace a prediction back to individual inputs and time scales. This lets us understand whether a model has captured physical relationships or artifacts of the data and can potentially tell us something new about the underlying physics itself.

In this talk, I will give a brief overview of some of these methods and how they can be used in space physics. I will then walk through several of them on one particular example: predicting electron fluxes measured by THEMIS in the Earth's plasma sheet, with solar wind parameters and geomagnetic indices as inputs. I will compare gradient boosting, a temporal convolutional network and a pretrained transformer model in this task. Finally, I will show how interpretability methods can help identify which drivers and time scales are important for the prediction.

Informationen

 

IWF Seminar series

Speaker
Alexander Lukin

When
19.11.2026, 13.45 Uhr

Where
U.a.4 in-person and via Zoom

Recordings
Please be aware that the talks may be recorded, including the questions asked by the audience after the talk.