Causality-Aware 3D/4D Geometry Learning for Scientific Discovery

Mahule Roy, Subhas Roy

Keywords: causal reasoning, counterfactual analysis, neural implicit representations, scientific computing, physics-informed learning

TL;DR: We introduce a causality-aware 3D/4D reconstruction framework that enables counterfactual reasoning on geometric evolution while maintaining high-quality scene reconstruction across scientific domains.

Abstract: Recent advances in 3D and 4D computer vision have enabled high-fidelity reconstruction of complex static and dynamic scenes from heterogeneous visual observations. Here, 3D refers to spatial geometry, while 4D captures the temporal evolution of these 3D shapes over time. However, most existing approaches remain fundamentally correlational, focusing on reproducing geometry and appearance without explicitly modeling the causal mechanisms that govern scientific phenomena. In many scientific domains—such as climate science, urban systems, and biomedicine—geometry is not merely an observable outcome but an active participant in underlying physical, biological, or environmental processes. We introduce a causality-aware framework for 3D/4D geometry learning that integrates causal reasoning, physical priors, and intervention-based analysis into neural reconstruction pipelines. Our approach enables counterfactual reasoning and “what-if” simulations directly on dynamic 3D scenes, while maintaining competitive reconstruction quality. Across glacier, urban flooding, and cardiac MRI datasets, we demonstrate modest but consistent improvements in generalization and counterfactual accuracy, and we carefully document limitations, computational requirements, and failure cases to provide a realistic assessment of capabilities.

Submission Number: 3