May 18 at 10 am
Doctoral Programme | Computer Science
Defense | Using Machine Learning to Refine Space Weather Models and Infer Its Effects on Earth
Student | Ana Filipa Sousa Barros
Date: May 18
Time: 10:00 am
Venue: Room FC5 003
President:
Miguel Tavares Coimbra
Full Professor
Faculty of Sciences, University of Porto
Examiners:
Sabrina Guastavino
Assistant Professor
University of Genova (Italy)
Ricardo Jorge Maranhas Gafeira
Director
Geophysical and Astronomical Observatory of the University of Coimbra
Committee Members:
Inês de Castro Dutra
Assistant Professor
Faculty of Sciences, University of Porto
André Monteiro de Oliveira Restivo (Orientador)
Associate Professor
Faculty of Engineering, University of Porto
Abstract:
The integration of space-based technologies into critical infrastructure, from Global Positioning System (GPS) to telecommunications satellites, has transformed space weather forecasting from an academic pursuit into an operational necessity. Modern society’s dependence on these systems creates unprecedented demands for accurate and timely solar-terrestrial predictions. Yet, current operational models face fundamental limitations that compromise their effectiveness during major events.
Traditional magnetohydrodynamic (MHD) simulations, though physically comprehensive,exhibit computational bottlenecks that preclude real-time operational deployment, often requiring hours to days to generate predictions that must be delivered within minutes for effective space weather operations. These issues are compounded by the integration of heterogeneous observational sources and the scarcity of in-situmeasurements for validation, leading to incomplete observational coverage of solar wind evolution.
This thesis investigates the hypothesis that:
H: Machine Learning (ML) models can improve both the accuracy and computational efficiency of solar wind predictions by learning complex non-linear relationships from historical solar and heliospheric data, outperforming traditional physics-based models in short to medium-term forecasting tasks, and supporting the understanding of solar wind interactions with Earth’s magnetosphere and their manifestations, such as auroral dynamics.
The research comprises six complementary studies: (1) neural optimisation of initial conditions for MHD simulations reduces the computational burden of expert-driven boundary condition selection; (2) clustering and adversarial anomaly detection improve predictive precision through model specialisation and data quality control; (3) physics-informed neural surrogates integrate conservation laws directly into network architectures, achieving rapid inference while maintaining physical consistency; (4) operational deployment within the SWiFT pipeline validates these methods under real forecasting conditions; (5) ML analysis of twenty-six years of European Incoherent Scatter Scientific Association (EISCAT) radar and OMNI data elucidates solar wind-magnetosphere coupling mechanisms; and (6) a multi-platform study examines the influence of Earth’s magnetic pole migration on auroral boundaries over two decades.
Empirical evaluation employs datasets spanning multiple solar cycles, including SDO magnetograms, coronal observations from multiple spacecraft, MULTI-VP simulation outputs, in-situ ACE and Deep Space Climate Observatory (DSCOVR) data at L1, and EISCAT groundbased radar measurements. Controlled experiments and comparative analyses ensure rigorous validation against operational baselines.
Results consistently support the central hypothesis. Neural optimisation achieved statistically significant reductions in computational cost. Physics-informed surrogates preserved MHD mass and momentum conservation with mean squared MSE errors below 5% relative to MULTI-VP solutions, while achieving inference times of 8 ms (physics-informed networks) and 50 ms(neural operators), being three to four orders of magnitude faster than conventional simulations. Operational tests within SWiFT confirmed robustness across solar conditions. Statistical performance evaluation revealed consistently high accuracy when validated against independent spacecraft observations, with physics-informed variants maintaining superior performance compared to baseline approaches across multiple evaluation criteria.
Long-term ionospheric analysis was done over four distinct Arctic regimes—auroral oval, cusp aurora, high electric field, and polar cap conditions, with effect sizes ranging from 0.3 to 0.9, demonstrating strong quantitative relationships between solar wind parameters and ionospheric behaviour. Solar wind temperature emerged as the dominant control factor for polar cap phenomena, while other auroral regions exhibited multi-parameter dependencies.
The study of magnetic pole migration revealed systematic auroral boundary shifts consistent with theoretical expectations: equatorward migration rates of 0.069◦ yr−1 (poleward boundary) and 0.041◦ yr−1 (equatorward boundary) in the Svalbard sector, validated by radar and magnetic field observations.
This research fundamentally challenges the conventional assumptions about trade-offs between computational efficiency and scientific accuracy, demonstrating that physics-informed ML can simultaneously enhance both dimensions when carefully designed and implemented. The findings of this thesis establish physics-informed neural architectures as tools for space weather forecasting, addressing critical limitations of previous approaches while providing computational advantages essential for operational deployment. The demonstrated ability to integrate MLefficiencywithphysicalrigour opensnewpossibilities for advancing both scientific understanding and practical forecasting capabilities, with implications extending beyond space weather to encompass numerous scientific computing applications requiring a similar balance between computational efficiency and physical accuracy.