Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Solaris: A Foundation Model of the Sun

Harris Abdul Majid, Pietro Sittoni, Francesco Tudisco

Abstract

Foundation models have demonstrated remarkable success across various scientific domains, motivating our exploration of their potential in solar physics. In this paper, we present Solaris, the first foundation model for forecasting the Sun's atmosphere. We leverage 13 years of full-disk, multi-wavelength solar imagery from the Solar Dynamics Observatory, spanning a complete solar cycle, to pre-train Solaris for 12-hour interval forecasting. Solaris is built on a large-scale 3D Swin Transformer architecture with 109 million parameters. We demonstrate Solaris' ability to generalize by fine-tuning on a low-data regime using a single wavelength (1700 Å), that was not included in pre-training, outperforming models trained from scratch on this specific wavelength. Our results indicate that Solaris can effectively capture the complex dynamics of the solar atmosphere and transform solar forecasting.

Solaris: A Foundation Model of the Sun

Abstract

Foundation models have demonstrated remarkable success across various scientific domains, motivating our exploration of their potential in solar physics. In this paper, we present Solaris, the first foundation model for forecasting the Sun's atmosphere. We leverage 13 years of full-disk, multi-wavelength solar imagery from the Solar Dynamics Observatory, spanning a complete solar cycle, to pre-train Solaris for 12-hour interval forecasting. Solaris is built on a large-scale 3D Swin Transformer architecture with 109 million parameters. We demonstrate Solaris' ability to generalize by fine-tuning on a low-data regime using a single wavelength (1700 Å), that was not included in pre-training, outperforming models trained from scratch on this specific wavelength. Our results indicate that Solaris can effectively capture the complex dynamics of the solar atmosphere and transform solar forecasting.

Paper Structure

This paper contains 16 sections, 2 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Model
  4. Encoder
  5. Processor
  6. Decoder
  7. Normalization and Transformation
  8. Configurations
  9. Experiments
  10. Pretraining
  11. Experimental setup.
  12. Results.
  13. Finetuning
  14. Experimental setup.
  15. Results.
  16. ...and 1 more sections

Figures (9)

  • Figure 1: Illustration of Solaris' encoder module. The encoder processes multi-wavelength solar observations from AIA, HMI, and EVE instruments through patch embedding, tokenization, perceiver-based aggregation using cross-attention, spatiotemporal encoding, and transformer stack processing. The encoded representation is then passed to the Swin Transformer (ViT) processor for temporal evolution modeling.
  • Figure 2: Multi-wavelength forecasting results from $\text{Solaris}_\text{T}$ for 2023-07-02 00:00 UTC. From left to right: input states $X^{t-1}$ and $X^t$, target state $X^{t+1}$, predicted state $\hat{X}^{t+1}$, and absolute error $|\hat{X}^{t+1} - X^{t+1}|$. Rows show results across different wavelengths (94, 131, 171, 193, 211, 304, 335, and 1600 Å).
  • Figure 3: Multi-wavelength forecasting results from $\text{Solaris}_\text{S}$ for 2023-07-02 00:00 UTC. From left to right: input states $X^{t-1}$ and $X^t$, target state $X^{t+1}$, predicted state $\hat{X}^{t+1}$, and absolute error $|\hat{X}^{t+1} - X^{t+1}|$. Rows show results across different wavelengths (94, 131, 171, 193, 211, 304, 335, and 1600 Å).
  • Figure 4: RMSE error during training of $\text{Solaris}_{\text{S}}$ and $\text{Solaris}_{\text{T}}$
  • Figure 5: RMSE error during finetuning of $\text{Solaris}_{\text{S}}$.
  • ...and 4 more figures