Physics-aware Causal Graph Network for Spatiotemporal Modeling

Sungyong Seo, Zijun Cui, Sam Griesemer, Joshua Hikida, Yan Liu

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: physics-informed deep learning; causal learning; spatiotemporal learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Interpretable physics equations are widely recognized as valuable inductive biases for constructing robust spatiotemporal models. To harness these valuable pieces of knowledge, existing approaches often presuppose access to the exact underlying equations. However, such an assumption usually doesn't hold, especially in the context of real-world observations. Conversely, causality systematically captures the fundamental causal relations across space and time that are intrinsically present in physics dynamics. Nevertheless, causality is often ignored as a means of integrating prior physics knowledge. In this work, we propose a novel approach that effectively captures and leverages causality to integrate physics equations into spatiotemporal models, without assuming access to precise physics principles.
Specifically, we introduce a physics-aware spatiotemporal causal graph network (P-stCGN). Causal relationships are analytically derived from prior physics knowledge and serve as physics-aware causality labels. A causal module is introduced to learn causal weights from spatially close and temporally past observations to current observations via semi-supervised learning. Given the learned causal structure, a forecasting module is introduced to perform predictions guided by the cause-effect relations. Extensive experiments on time series data show that our semi-supervised causal learning approach is robust with noisy and limited data. Furthermore, our evaluations on real-world graph signals demonstrate superior forecasting performance, achieved by utilizing prior physics knowledge from a causal perspective.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6312