Toward Explainable AI for SKA-Scale Solar Radio Astronomy
MSE | Research projects in the MRU Computer Science – Results
Advisors
Prof. Dr. Brandon Panos, HSI I4DS
In colabroation with the Indian Institute of Technology, Kanpur through Prof. Dr. Rohit Sharma
More to explore: Square Kilometer Array SKAO
Project Abstract
Future solar radio observations, such as those expected from the Square Kilometre Array (SKA), will contain complex spatial, temporal, and spectral information. This project investigates mutual-information-based methods for identifying dependencies in such data and localizing which regions of one variable are informative for another. Generative models, particularly diffusion models, are increasingly used in astrophysics, yet their outputs are rarely explained. This project aims to move beyond generating visually realistic images by developing methods that reveal scientifically meaningful patterns and relationships within the data. As a first step toward explainable generative modelling for solar radio astronomy, the project applies neural mutual-information estimation to solar time series and develops diffusion-based information maps that identify which regions of a conditioning radio image contribute to reconstructing a target image.
Figure 1: Main architectral idea for calculating MI for 1D data using a Mutual-Information Neural Estimator (MINE)-network.
Figure 2: Validation MI curves for the pair totpot and TOTUSJZ for HARPNUM 11786 using different window sizes.
Above: results for the MINE-network as a function of training time and hyperparameter setting. A lower bound to the true degree of dependency between two random variables can be calculated by training the model until it converges and stable sizes, here around a value of 2.
Figure 3: UMAP visualization of the preprocessed radio-image distribution after inicial PCA dimensionality reduction.
Above: 2D embedding of simulated radio data from our IIT colleagues. Different colors represent different frequency bands. The projection shows how complicated our radio solar data it.
Figure 4: Training pipeline of the conditional diffusion model. A target image Xo is noised at a randomly sampled timestep t to obtain X1. The noisy image X1 is concatenated with the checkerboard condition y and passed to a lightweight U-Net. The model is additionally conditionned on the timestep t and frequency label f, and is trained to predict the injected noise €. Here, t is 300 and f is 196 MHz, labelled as 5.
Above: Extending the mutual information calculation to 2D and discommodes. We create a synthetic controlled experiment to judge the performance of different model explainability techniques.
Figure 5: Convergence of MI maps with respect to the number of steps N and trajectories/noise samples K. The bands indicate the standard deviation across four different samples.
Above: The two tested methods converge at different raitas to different end point.
Figure 6: Example fake images, Grad-CAM maps thresholded at CAM ≥0.7, fake hotspot regions, and sampled real counterpart images.
Above: Although the diffusion model creates convincing fake images, a simple CNN as a judge can tell the synthetic from the generated apart with high reliability. A general flaw with generative models.
Figure 7: Two-dimensional UMAP projection of the four image statistics q05, masked_fraction, std, and mean. The real and fake samples are clearly separated.
Training a linear model to classify classes based on select statistical features of the two classes is possible. A dimensionality reduction and embedding reveals a clear linier separability between the two classes, even though to a human the generated and synthetic images look identical.