Title: PnP-Corrector: A Universal Correction Framework for Coupled Spatiotemporal Forecasting

Abstract: Accurately predicting the future evolution of multiple interacting dynamical systems, such as those found in climate models, relies heavily on coupled spatiotemporal forecasting. However, current methodologies are hindered by a persistent bottleneck: compounding errors. In coupled systems, biases originating from individual subsystem simulators propagate and amplify each other—a phenomenon we define as Reciprocal Error Amplification—which causes long-range predictions to deteriorate rapidly. To overcome this obstacle, we introduce PnP-Corrector (Plug-and-Play Corrector), a universal correction framework. The central premise of this approach is to separate physical simulation from error correction: it keeps pre-trained physics simulation engines frozen while training a dedicated correction agent to proactively neutralize the systematic biases inherent to the coupled system. Additionally, we developed DSLCast, an efficient predictive model architecture that serves as the backbone for this framework. Comprehensive experiments indicate that our approach markedly improves both the long-term stability and accuracy of coupled forecasting systems. Notably, in the demanding task of a 300-day global ocean-atmosphere coupled forecast, PnP-Corrector decreased the baseline model’s prediction error by 28% and outperformed state-of-the-art models across several key metrics.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC