Observing and forecasting OH airglow movement in the night sky is crucial for understanding atmospheric dynamics. To estimate physical hyperparameters (such as wavelength, amplitude, propagation angle, and time-period), we utilize simulated wave image sequences where these wave dynamics are highly controlled. However, atmospheric conditions are rarely simple; reality involves complex interference patterns of superimposed waves. It remains an open question how neural networks adapt their internal representations when the physical complexity of the sequence increases.

This project investigates the evolution of CNN filters as they scale from simple to complex physical phenomena. The student will analyze how filters evolve when multiple simulated waves are superimposed on top of each other, creating interference. By visualizing the convolutional filters and utilizing representation similarity metrics like Centered Kernel Alignment (CKA), the goal is to analyze whether the network learns completely new feature extractors to handle interference patterns, or if it simply recombines the elementary wave filters in deeper layers.

The scope of this topic can be adjusted for a Bachelor Thesis, Master Thesis, or a Research Project (Forschungsmodul) based on the required credits.

Literature:

[1] Cranmer, M. et al. "The frontier of simulation-based inference." PNAS (2020): https://arxiv.org/abs/1911.01429

[2] Kornith, S. et al. "Similarity of Neural Network Representations Revisited." PMLR (2019): https://arxiv.org/abs/1905.00414

[4] Boelts et al., (2025). sbi reloaded: a toolkit for simulation-based inference workflows. Journal of Open Source Software, 10(108), 7754,  https://doi.org/10.21105/joss.07754

If interested, please contact Pratik Mayekar (pratik.mayekar@uni-a.de).

In atmospheric physics and spatio-temporal forecasting, deep learning models are used in predicting physical hyperparameters (e.g., wavelength, amplitude, angle, time-period) from simulated sequences of OH airglow waves. While these networks are effective at regression tasks, they often act as "black boxes." When a network predicts multiple physical parameters simultaneously from a single superimposed wave image, it is unclear what spatial and temporal features it actually utilizes for each specific parameter. Understanding this is vital to ensure the network is learning true physical properties rather than exploiting statistical artifacts in the simulation data.

This project focuses on adapting Explainable AI (XAI) techniques, which are traditionally used for classification, to a multi-output regression context. Attribution methods such as Grad-CAM [1] or Integrated Gradients [2] will be applied to a network trained on simulated wave images with controlled dynamics. The primary objective is to investigate whether the network spatially disentangles the physics—for example, evaluating if the model specifically focuses on the distance between wave crests to determine the wavelength, while analyzing the global orientation of the image to determine the propagation angle.

The scope of this topic can be adjusted for a Bachelor Thesis, Master Thesis, or a Research Project (Forschungsmodul) based on the required credits.

Literature: 

[1] Selvaraju, R. R. et al. "Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization." ICCV (2017): https://arxiv.org/abs/1610.02391 

[2] Sundararajan, M. et al. "Axiomatic Attribution for Deep Networks." PMLR (2017): https://arxiv.org/abs/1703.01365

If interested, please contact Pratik Mayekar (pratik.mayekar@uni-a.de).