Speaker
Description
Machine‑learning techniques, whether supervised, self‑supervised, or unsupervised, have become indispensable tools in the modelling of space weather in recent years. By tapping into the vast, heterogeneous archives collected over decades, they produce prediction models that are both swift and highly accurate, often rivalling or even surpassing traditional physics‑based approaches, though sometimes at the cost of interpretability.
The presentation will centre on SERENADE, a deep‑learning pipeline that predicts the daily maximum Kp index several days in advance from SDO/AIA EUV images at 193 Å. Initial prototypes already achieved performance comparable to the state of the art for fast‑solar‑wind‑driven events, yet several limitations were identified. Recent work has replaced the generic GoogLeNet latent‑vector extractor with a variational auto‑encoder specifically trained on solar images, yielding a more physically‑meaningful latent space and more stable forecasts.
A major thrust of the current effort is the enrichment of the training database. Beyond the original SDOML collection (2010‑2020), we have assembled a new ML‑prepared dataset that spans 2010‑mid‑2025 (including the rising and maximum phases of the present, more active cycle) and a complementary SOHO/EIT dataset extending back to 1996, thereby providing nearly 30 years of observations. By analysing the impact of this temporal extension, we explore cross‑instrument generalisation (zero‑shot learning) by training on one instrument’s data and testing on the other, a capability that will become crucial once the SDO and SOHO missions are retired.
This work highlights critical aspects of data preparation and model assessment, and demonstrate how expanding and harmonising large data archives can improve the reliability and operational readiness of space‑weather forecasts.
| Numerical model | SERENADE |
|---|
