Speaker
Description
In recent years, studies have shown that it is possible to predict the geoeffectiveness of solar activity (L1 solar wind speed, geomagnetic indices) directly from EUV solar images using deep learning models (Upendran et al. 2020; Bernoux et al. 2022; Brown et al. 2022; Hu et al. 2022; Wang et al. 2025). These models, which are ultimately intended to be used operationally to provide early warnings of space weather events, are currently at the prototype or proof-of-concept stage and, although already more accurate than most other approaches, have numerous limitations. This is the case for the SERENADE model (Bernoux et al. 2022), which predicts the daily maximum of the Kp index a few days in advance from EUV images at 193A provided by the Atmospheric Imaging Assembly (AIA) instrument on the Solar Dynamics Observatory mission. Although the prediction performance for fast solar wind driven events was at least as good as the current state of the art, several weaknesses in the model were identified. The first weakness was that a latent vector was extracted from each image using a pre-trained GoogLeNet model instead of a model specific to solar images. This was addressed by Tahtouh et al. (under revision at JGR: MLC), who show that a Variational AutoEncoder results in a better structured latent space that correlates better with the physical properties of the Sun and leads to more stable and credible predictions.
However, the gain in accuracy with SERENADE was modest. We are exploring another avenue here, which lies in the training dataset used. So far, we've only used images from the SDOML dataset with its limited temporal coverage (2010 - 2020) during a weak solar cycle. We now generated our own ML-prepared dataset of SDO/AIA images from 2010 to mid-2025, adding 4.5 years of data during the rising and maximum phases of the current, much more active solar cycle. In addition, we use a dataset of SOHO/EIT images, also ML-prepared, which allows us to extend our dataset back to 1996, benefiting from nearly 30 years of data instead of the original 11. We analyze the benefits of this temporal extension and assess the extent to which our previous model may have been underestimated. Given that many studies also rely exclusively on the use of the SDOML dataset, our results are potentially generalizable to other models, and may indicate whether performance gains could be achieved without changing the architecture but simply increasing the database.
In addition, we take advantage of having two datasets from different instruments to perform a preliminary study of the extent to which training such a model with one dataset can produce usable results when used with data from another instrument (zero-shot learning), which would be of interest in preparing for the future when the SOHO and SDO missions have been discontinued.
| Do you plan to attend in-person or online? | In-person |
|---|
