DSTGN: Decoupled SpatioTemporal Graph Network for Multi-scenario Dynamics Learning

Yuan Mi, Qi Wang, Han Wan, Xueqin Hu, Yang Liu, Hao Sun

Keywords: decoupled modeling, dynamic prediction, multi-scenario learning, graph network

Abstract: Machine Learning (ML) methods have played a pivotal role in a wide range of dynamics tasks, such as physical simulation, multi-modal interaction, and real-time prediction. Among them, Graph Neural Networks (GNNs) have emerged as effective surrogates for numerical solvers, thanks to their ability to model pairwise interactions between nodes and their neighboring edges. However, the inherent tendency of GNNs to perform local aggregation limits their expressiveness in capturing intricate global interactions. To address this limitation, we propose a general framework, Decoupled SpatioTemporal Graph Network (DSTGN), to enhance prediction performance across various downstream tasks. DSTGN enhances latent feature representations by decoupling spatial connectivity in both the physical and latent spaces. In parallel, it introduces a learnable temporal integration mechanism that decouples inter-step dynamics at each latent layer, effectively mitigating error accumulation during autoregressive inference. Extensive experiments on multiple types of benchmarks demonstrate that DSTGN consistently outperforms existing baselines in both accuracy and generalization, across regular and irregular domains, as well as static and dynamic meshes.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 2563