Dingling Yao

This PhD project focuses on developing causal representation learning algorithms to construct digital twins of scientific systems with explicitly identified causal mechanisms. Directly deriving scientific insights from real-world experimental data is challenging for multiple reasons, primarily due to the multimodal nature and inherent noise in the collected data. A prime example is climate modeling, where intricate interactions between atmospheric, oceanic, and terrestrial systems generate vast amounts of noisy, high-dimensional measurements. At the same time, many scientific questions are inherently causal, as all natural measurements we record are governed by physical laws. Understanding these laws requires a high-level abstraction of experimental observations that preserves true causal latent factors and their relationships. By uniquely identifying these causal mechanisms, this project provides a principled framework for modeling underlying physical processes and answering downstream causal queries across various scientific domains.