Skip to content

Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction

Home / Papers / Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction

Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction

Authors: Xuanle Zhao, Yue Sun, Ziyi Wang, Bo Xu, Tielin Zhang Date: 2026-04-21 Paper ID: openalex:2405.14504

Summary

This paper addresses the limitations of existing physics-informed neural networks in capturing complex spatiotemporal dynamics by introducing an adaptive second-order Runge-Kutta integration scheme. By embedding this adaptive solver directly into the architecture, the model improves the precision of state-transition estimations for PDEs. Additionally, a frequency-enhanced Fourier module is utilized to capture complex dynamic patterns, resulting in superior performance on spatiotemporal and video prediction tasks with higher computational efficiency.

Key Contributions

  • Introduces a physical-guided neural network architecture that integrates an adaptive second-order Runge-Kutta solver to accurately track spatiotemporal state evolution.
  • Proposes a frequency-enhanced Fourier module to improve the estimation of spatiotemporal dynamics, capturing complex dependencies better than standard convolutional or recurrent alternatives.
  • Achieves state-of-the-art performance on spatiotemporal and video prediction benchmarks while significantly reducing total parameter count.

Open Questions & Future Work

Key Concepts

  • adaptive-runge-kutta-dynamics: A neural network integration method that employs an adaptive second-order Runge-Kutta solver to explicitly model physical state transitions in spatiotemporal dynamics.

Archivist Review

I approved the adaptive Runge-Kutta concept as a distinct methodological contribution for differentiable physics and the open question regarding PDE learning generalization as a foundational research bottleneck. The frequency-enhanced module was rejected as it represents a common architectural pattern rather than a novel contribution. No datasets were approved as none were specifically named or defined as the central reusable benchmark for the field.

Approved Concepts

  • Adaptive Runge-Kutta Dynamics: It shifts the paradigm from fixed-step neural integration to adaptive solving, providing a more robust architectural primitive for physics-informed spatiotemporal modeling.

Approved Open Questions

  • Generalizing PDE Learning Methods: Addressing this bottleneck is essential for the move from task-specific PINNs to general-purpose foundation models for physical dynamics.

Rejected Candidates

  • [concept] Frequency-enhanced Fourier Module (frequency-enhanced-fourier-module) - not_novel: Fourier-based modules are already well-established; this specific variant is a minor architectural variation rather than a fundamental innovation.

Metadata & Links

url
https://arxiv.org/abs/2405.14504
paper_id
2405.14504
paper_source
openalex
domain
time-series
tags
time-seriesspatiotemporal-predictionphysics-informed-learningpde-modelingvideo-prediction
architectures
datasets
concept_slugs
adaptive-runge-kutta-dynamics
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-24T05:51:00Z