Diffusion model generates synthetic images of the Sun sorted by flare strength

The work

Technical · from the record

The study trains denoising diffusion probabilistic models to generate synthetic 64x64 full-disc images of the Sun in the SDO/AIA 171 Å channel, conditioned on the GOES flare class of the image; HEK flare peak times were matched to GOES X-ray flux and to AIA frames to build a set of 20,420 labelled images. Three conditioning strategies were compared — discrete GOES classes, continuous GOES X-ray values, and discrete labels combined with ceVAE latent features — and 60,000 generated images were assessed with cluster metrics, FID and a classifier-based F1 score; the discrete-label model reached a macro F1 of 0.38 against 0.54 for the same classifier applied to true data. Adding generated images of the under-represented classes to a supervised classifier's training set raised per-class accuracy, for example 81.9% for the A class with 600 added images compared with 30.1% with none, and 29.1% for the M class compared with 7.1%, while classical image augmentations did not consistently raise accuracy. In a binary 24-hour flare-prediction experiment on 2282 daily images, training with DDPM-generated data gave a TSS of 0.35 ± 0.02 and an HSS of 0.18 ± 0.02.

How AI was used

A denoising diffusion probabilistic model with a U-Net backbone containing self-attention layers between downsampling and upsampling blocks was trained from scratch on AIA 171 Å images at 64x64 resolution, using AdamW, an MSE loss, a learning rate of 3e-4, a batch size of 12 and 500 epochs on one NVIDIA TITAN X GPU, with PyTorch. Images were labelled by cross-matching HEK flare peak times with GOES XRSB flux and then with the nearest subsequent SDOMLv2 frame within a 7-minute tolerance. Three conditioned variants were trained using classifier-free guidance: an embedding layer over the discrete GOES classes, a two-layer encoding of the continuous X-ray value summed with the time-step embedding, and the discrete label concatenated with latent features from a ceVAE previously pretrained on SDO data. Sampling was then run with a specified flare-class label to produce synthetic images in both uniform and dataset-matched class proportions. Evaluation used K-means cluster metrics in the ceVAE latent space, FID computed in CLIP ViT-B/32 and InceptionV3 feature spaces with the clean-fid library, t-SNE for inspecting candidate feature spaces, and a supervised DeIT classifier trained on real images and applied to generated images. The synthetic images were then used as training data in two further experiments: a multi-class DeIT classifier with 200, 400 or 600 generated images added to the A, M and X classes and compared against torchvision geometric augmentations, and a binary 24-hour flare predictor on one image per day with a weighted cross-entropy loss, cosine learning-rate decay, 18 epochs, and 50 to 500 generated images injected per under-represented class.