Spatio-temporal Twins with A Cache for Modeling Long-term System Dynamics

Hao Wu, Kun Wang, Fan Xu, Yue Li, Xu Wang, Weiyan Wang, Haixin Wang, Xiao Luo

Primary Area: applications to physical sciences (physics, chemistry, biology, 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: long-term dynamical systems, spatio-temporal forecasting

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

Abstract: This paper investigates the problem of modeling long-term dynamical systems, which are essential for comprehending fluid dynamics and astrophysics. Recently, a variety of spatio-temporal forecasting approaches have been proposed, which usually employ complicated architectures (e.g., Transformer) to learn spatial and temporal relationships. However, these approaches typically perform poorly for long-term forecasting due to information loss during exploration and iterative rollouts. To tackle this, we propose a new framework named Spatio-temporal Twins with A Cache (STAC) for long-term system dynamics modeling. To investigate spatio-temporal relationships from complementary perspectives, STAC contains a frequency-enhanced spatial module and an ODE-enhanced temporal module. Then, we fuse the information between twin modules with channel attention for discriminative feature maps. To capture long-term dynamics, we introduce a cache-based recursive propagator, which stores the previous feature maps in the cache memory during recursive updating. Moreover, we involve both teacher forcing with Mixup and semi-supervised adversarial learning to enhance the optimization process. Extensive experiments show that the proposed STAC can achieve superior performance to existing state-of-the-art methods.

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: 2441