Emam Hossain successfuly defends his PhD Proposal!

Causal reasoning is fundamental to understanding dynamic processes in complex scientific systems such as climate, glaciology, and remote sensing. While modern deep learning models have demonstrated strong performance on sequence modeling tasks, they often rely on purely correlational representations, limiting their generalization and interpretability under distribution shifts or partial observability. This dissertation presents a principled progression of methods that integrate causal discovery and representation learning for modeling dynamic systems. First, we reinterpret Reconstructed Phase Space (RPS) modeling as an unsupervised representation learning approach for satellite-derived time series, enabling interpretable classification of supraglacial lake evolution via Gaussian Mixture Models. We then extend this with a causally-informed temporal modeling framework using PCMCI+ to identify stable, regionally meaningful predictors, which improve generalization and robustness in sequence classification models. However, both methods operate in the space of observed variables and do not recover unobserved mechanisms. To address this limitation, we propose a novel Causal Representation Learning (CRL) framework for dynamic systems that employs variational autoencoders with interventional regularization to learn low-dimensional latent variables aligned with underlying causal factors. This framework is designed to support do-operations and counterfactual reasoning in latent space, and will be validated on benchmark datasets with known causal graphs (Pendulum, Flow, CelebA), before being applied to supraglacial lake dynamics as a real-world use case. Collectively, this work aims to advance the boundary of causal modeling in temporal systems by bridging the gap between statistical representation learning and structural causal inference, offering scalable and interpretable tools for scientific discovery in dynamic, high-dimensional environments.

• More Information about Emam Hossain successfuly defends his PhD Proposal!