Deep Generative model that uses physical quantities to generate and retrieve solar magnetic active regions

TL;DR

The paper tackles the interpretability gap in deep generative models for scientific data by coupling GANs with SVMs to map latent space directions to physical solar AR parameters ($TUF$, $PSEP$, $R$) and using SSL to retrieve real matches. A GAN generates patches from a $100$-dimensional latent vector $z$ and is steered along latent directions corresponding to $TUF$, $PSEP$, and $R$ via SVMs. SimSiam SSL is used to map generated queries to a latent space of real SHARP patches, enabling retrieval of matches with statistically meaningful agreement in $TUF$, $R$, $PSEP$, $TPF$, and $TNF$. The results show that latent directions produce interpretable changes and the SSL-based retrieval yields real patches that closely match the generated queries, with reported correlations such as Pearson ≈ 0.78 and Spearman ≈ 0.79 for $TUF$. The work demonstrates a pathway to use generative models as scientific interrogators with public data and code, and can extend to other astronomical datasets.

Abstract

Deep generative models have shown immense potential in generating unseen data that has properties of real data. These models learn complex data-generating distributions starting from a smaller set of latent dimensions. However, generative models have encountered great skepticism in scientific domains due to the disconnection between generative latent vectors and scientifically relevant quantities. In this study, we integrate three types of machine learning models to generate solar magnetic patches in a physically interpretable manner and use those as a query to find matching patches in real observations. We use the magnetic field measurements from Space-weather HMI Active Region Patches (SHARPs) to train a Generative Adversarial Network (GAN). We connect the physical properties of GAN-generated images with their latent vectors to train Support Vector Machines (SVMs) that do mapping between physical and latent spaces. These produce directions in the GAN latent space along which known physical parameters of the SHARPs change. We train a self-supervised learner (SSL) to make queries with generated images and find matches from real data. We find that the GAN-SVM combination enables users to produce high-quality patches that change smoothly only with a prescribed physical quantity, making generative models physically interpretable. We also show that GAN outputs can be used to retrieve real data that shares the same physical properties as the generated query. This elevates Generative Artificial Intelligence (AI) from a means-to-produce artificial data to a novel tool for scientific data interrogation, supporting its applicability beyond the domain of heliophysics.

Figures (6)

• Figure 1: Graphical illustration of the designed and implemented solar image generation and retrieval pipeline. The process involves a generative model (GAN), supervised learning to utilize physical parameter space, and self-supervised learning (SSL) for reverse real image-search. The Look-Up-Table (LUT) describes the mapping between an image and SSL-derived latent representation.
• Figure 2: A zoo of GAN-generated magnetic Active Regions that depending on complexity play an important role in driving space weather events. The shown AR images are generated using 100 latent vectors randomly sampled from a standard normal distribution. The images with diffused flux are seen as the temporal evolution of SHARPs is used during the training of the GAN.
• Figure 3: Calculation of physical parameters from thresholded GAN generated magnetic patches. The top row shows a GAN-generated active region, its thresholded version, and the derived regions of polarity inversion with a high field gradient from left to right. The collage starting from the second row shows the thresholded version of all GAN-generated images in Figure \ref{['fig2']} with field-strength-weighted centroids of positive and negative polarity connected by a cyan line.
• Figure 4: Generated solar images overlaid on the 2-dimensional projection of latent space. The projection was performed such that the decision boundary hyperplane becomes a line (marked in red) in the depicted 2d latent space. The decision boundary marks a clear separation between the physical properties (total unsigned field (TUF), and polarity) represented by the generated images. The bottom rows show zoomed-in views (outlined with green and blue rectangles) of the generated images from both extremes across the decision boundaries.
• Figure 5: Manipulation of GAN generated image along two different physically interpretable directions. Two directions are shown -- total unsigned field (TUF; top two rows) and polarity separation (PSEP; bottom three rows). The 2nd row from the top shows the result of decoupling TUF from the direction along which the R changes. The 4th row shows the result of decoupling PSEP from the direction along which polarity is flipped and the bottom-most row shows the result of decoupling PSEP from both polarity flip and change in R.