Core Concepts
ST-PAD framework enhances fluid dynamics modeling through physical-awareness and parameter diffusion guidance.
Abstract
The paper introduces the ST-PAD framework for spatio-temporal fluid dynamics modeling. It consists of an upstream stage focusing on physical constraints and a downstream stage utilizing parameter diffusion. Extensive experiments validate its effectiveness in outperforming mainstream models, especially in capturing local representations and maintaining advantages in OOD generations.
Upstream Stage:
Design of vector quantization reconstruction module with temporal evolution characteristics.
Introduction of general physical constraints for balanced parameter distribution.
Downstream Stage:
Utilization of diffusion probability network involving parameters for high-quality future states generation.
Enhancement of model's generalization ability across various physical setups.
Contributions:
Proposal of ST-PAD framework for fluid dynamics modeling.
Systematic locking of task rules in upstream stage and fine-tuning model's generalization ability in downstream stage.
Demonstrated strong generalization capabilities through experiments on benchmark datasets.
Stats
Extensive experiments on multiple benchmark datasets have verified the effectiveness and robustness of the ST-PAD framework, showcasing that ST-PAD outperforms current mainstream models in fluid dynamics modeling and prediction.