Course Overview
8th to 12th June 2026 – KTH, Stockholm, Sweden
This one-week summer school introduces modern machine learning (ML) methods applied to heliophysics and magnetospheric science, aimed at graduate students and early-career researchers. The program combines foundational lectures with hands-on coding sessions, enabling participants to build both conceptual understanding and practical skills.
Participants will explore core ML techniques—including supervised and unsupervised learning methods—and learn how to access, preprocess, and analyze real datasets from major space missions. Guided by experienced mentors, attendees will work directly with mission data, develop reproducible workflows, and implement ML models relevant to current research challenges in space physics.
The summer school emphasizes interactive learning, collaborative problem-solving, and discussion of best practices in scientific machine learning. At the end of the week, participants will have the opportunity to present their work or research interests in short lightning talks presentations, providing a platform for feedback, discussion, and networking.
Please note that the Summer School does not offer fellowships or financial support. Participants are responsible for their own travel and accommodation costs. The summer school does not include lunch or dinner. Participants are expected to arrange and pay for their own meals.
🎯Learning Objectives
By participating in the Summer School, attendees will be able to:
- Explain core concepts in space physics and identify key machine-learning applications in the field.
- Locate, access, and preprocess datasets from major missions such as Magnetospheric Multiscale Mission, Cluster, THEMIS, Solar and Heliospheric Observatory, Solar Dynamics Observatory, Solar Orbiter, and OMNI, applying workflows for cleaning, normalization, feature extraction, and time alignment.
- Apply and interpret dimensionality-reduction and unsupervised techniques (e.g., PCA, clustering, covariance analysis) to uncover structure and regimes in space-physics data.
- Design, implement, and validate machine-learning models for classification, prediction, and event detection in space-physics applications.
- Critically assess model performance, limitations, uncertainty, and physical interpretability, and discuss the role of ML in current and future space missions.
- Demonstrate proficiency with Python-based scientific and ML tools, such as scikit-learn, Pandas, and PyTorch.
Prerequisites
Basic Python programming and introductory linear algebra are recommended. No prior machine-learning expertise is required.
Who should apply:
This summer school is designed for graduate students and early-career researchers who are studying or conducting research in space physics, space weather, or related areas. As well as those with a machine learning background who are beginning to apply their skills in these fields.
Participation in the summer school is free of charge, but registration is mandatory and subject to approval by the organizer.
Sign-up using the form: https://forms.gle/6n79Vdw2Nxxm3BPF9
The sign-up will close on the 30th of April, 2026.
Credit: ESA/ATG medialab
Instructors and Organizers
Jonah Ekelund (Co-Organizing Chair, Instructor)
KTH Royal Institute of Technology
jonahek (at) kth.se
Stefano Makidis (Co-Organizing Chair, Instructor)
KTH Royal Institute of Technology
George Miloshevich (Instructor)
KU Leuven
Ekaterina Dineva (Instructor)
KU Leuven
Panagiotis Gonidakis (Instructor)
KU Leuven
Edoardo Legnaro (Instructor)
University of Genova
Federica Bragone (Instructor)
KTH Royal Institute of Technology
Course Materials
Welcome and Introduction
Lecture recording: https://youtu.be/nWbqMfjsFUo
ML Fundamentals
This lecture introduces the foundations of machine learning for participants with little or no prior ML background. Starting from first principles, it covers the relationship between AI, machine learning, and deep learning, and builds up the core components common to any ML system: data, features, model, loss function, optimization, and evaluation. The lecture surveys the three main learning paradigms — supervised, unsupervised, and reinforcement learning — and introduces key algorithms including k-NN, SVM, k-means, and neural networks. It also touches on modern architectures (CNNs, RNNs, Transformers) and emerging directions in scientific ML and generative AI. By the end of the session, participants will have the conceptual vocabulary and mental framework needed to engage with the more applied, space-physics-focused material throughout the rest of the week.
Presentation: See further down.
Light introduction to Python: https://github.com/Jonah-E/magnetosphere_classification/blob/main/00_prerequisites.ipynb
Magnetospheric Classification
This module covers the full pipeline from raw spacecraft data to a working classifier, using the problem of automatically identifying plasma regions in MMS ion spectrograms as the running example. It begins with where magnetospheric data comes from — with a close look at the Cluster and MMS missions — and the practical routes for accessing it in Python via tools such as pyspedas, cdasws, and the ASAP-developed SpacePhyML library. From there, it addresses the less glamorous but critical steps that follow: inspecting a dataset for missing values, time gaps, and unexpected artifacts; understanding what your features actually represent; and making deliberate, physics-informed choices about how to clean and represent the data before it reaches a model. These foundations feed directly into the modeling pipeline — log-scaling, PCA-based dimensionality reduction, unsupervised clustering (k-means, GMM, agglomerative), and supervised classification with a convolutional neural network in PyTorch — along the way covering class imbalance, data leakage, train/validation/test splits, and how to read a confusion matrix.
Lecture recording:
Lecture slides: https://github.com/Jonah-E/magnetosphere_classification/tree/main/presentations
Lecture notebooks: https://github.com/Jonah-E/magnetosphere_classification/tree/main
Exercises: https://github.com/Jonah-E/magnetosphere_classification/tree/main/exercises
Clean labeled data: https://kth-my.sharepoint.com/:u:/g/personal/jonahek_ug_kth_se/IQCY30_7U8NUQZHM1sWQp-vhAX0hUq3TnaJvbQGrze2Z56A?e=amNgDx
Space weather and Deep Learning tools
This module builds deep learning from first principles — perceptrons through Transformers — using solar flare forecasting as the application thread throughout. The physics motivation is established first: how active regions, solar flares, and CMEs drive space weather and why reliable forecasting matters. Core deep learning architectures (MLPs, CNNs, ResNets, RNNs, Transformers) are then introduced progressively, each motivated by the limitations of the previous one, alongside practical training techniques such as data augmentation, dropout, and the Adam optimizer. The second half demonstrates a complete real-world pipeline from the ARCAFF project: building a multi-decade magnetogram dataset, classifying active regions by magnetic complexity using Vision Transformers, and forecasting flares 24 hours ahead with a hybrid CNN–Transformer model trained on multi-channel SDO/HMI time series. The module closes with a concise introduction to MLOps and reproducibility practices.
Lecture recording:
Lecture slides: https://github.com/edoardolegnaro/KTH_ML4Space2026/blob/main/slides/2026-06-10_ASAP_KTH_SummerSchool.pdf
Exercises: https://github.com/edoardolegnaro/KTH_ML4Space2026/blob/main/Solar_AR_cutouts.ipynb
Machine Learning for Heliophysics
This module introduces machine learning techniques applied to solar and heliospheric science. Following a lecture on solar physics fundamentals — covering activity cycles, the solar wind, and coronal mass ejections — and a practical overview of key datasets from SOHO, SDO, Solar Orbiter, and OMNI, the module covers two main methodological areas. The first focuses on solar image segmentation and detection, working through both classical computer vision approaches — including multi-level Otsu thresholding and morphological filtering — and modern deep learning architectures such as U-Net and YOLOv8, applied to SDO/AIA observations. Methods are compared in terms of accuracy, computational cost, and suitability for onboard spacecraft deployment, complemented by a hands-on component combining segmentation, feature extraction, and clustering. The second area introduces Gaussian Process Regression as a Bayesian framework for probabilistic forecasting, with applications to solar wind prediction and CME Earth-arrival time estimation using the HUXt heliospheric model. Participants apply GPR to predict CME hit probabilities and arrival times from cone model parameters, building intuition for surrogate modelling of computationally expensive simulations and for quantifying forecast uncertainty.
Lecture Recordings:
- https://youtu.be/8V0kUsCOokg
- https://youtu.be/ZPqD-oWBuvc
- https://youtu.be/_Ol8zoBEjAE
- https://youtu.be/CaR0ubazp2Q
Lecture slides:
- https://github.com/pangonidakis/ASAPSummerSchool4SpaceMLHandsOn/blob/main/ASAPWorkshop2026_KULeuven.pdf
- https://github.com/georgemilosh/conecast/blob/main/ASAPWorkshop2026_KULeuven.pptx
Exercises:
Sponsors
