PDE-regularized Dynamics-informed Diffusion with Uncertainty-aware Filtering for Long-Horizon Dynamics
Min Young Baeg, Yoon-Yeong Kim · 2026
Read PDE-regularized Dynamics-informed Diffusion with Uncertainty-aware Filtering for Long-Horizon Dynamics by Min Young Baeg, Yoon-Yeong Kim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09058.txt.
Introduction
Baeg and Kim tackle long-range guessing, where piling errors, noise growth, and missing real-world fit defeat current models. Diffusion models offer a chance-based frame for doubt but usually rely on mean squared error and miss motion ruled by nature's laws. Their PDYffusion system brings PDE-based controls into a motion-aware model with doubt filter. It settles under the new loss. Tests on several data sets show better CRPS and MSE scores with stable doubt, giving an even, robust answer to the score-doubt trade-off.
Catalog record
| CREATOR | Min Young Baeg, Yoon-Yeong Kim |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68033 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09058
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PDE-regularized Dynamics-informed Diffusion with Uncertainty-aware Filtering for Long-Horizon Dynamics Min Young Baeg, Yoon-Yeong Kim Abstract Long-horizon spatiotemporal prediction remains a challenging problem due to cumulative errors, noise amplification, and the lack of physical consistency in existing models. While diffusion models provide a probabilistic framework for modeling uncertainty, conventional approaches often rely on mean squared error objectives and fail to capture the underlying dynamics governed by physical laws. In this work, we propose PDYffusion, a dynamics-informed diffusion framework that integrates PDE-based regularization and uncertainty-aware forecasting for stable long-term prediction. The proposed method consists of two key components: a PDE-regularized interpolator and a UKF-based forecaster. The interpolator incorporates a differential operator to enforce physically consistent intermediate states, while the forecaster leverages the Unscented Kalman Filter to explicitly model uncertainty and mitigate error accumulation during iterative prediction. We provide theoretical analyses showing that the proposed interpolator satisfies PDE-constrained smoothness properties, and that the forecaster converges under the proposed loss formulation. Extensive experiments on multiple dynamical datasets demonstrate that PDYffusion achieves superior performance in terms of CRPS and MSE, while maintaining stable uncertainty behavior measured by SSR. We further analyze the inherent trade-off between prediction accuracy and uncertainty, showing that our method provides a balanced and robust solution for long-horizon forecasting. 1 Introduction In many applied scientific domains, accurately tracking and forecasting dynamical systems over long time horizons remains a fundamental and challenging problem. Traditional numerical approaches grounded in explicit mathematical modeling often incur substantial computational cost and become impractical for large-scale or long-term simulations. As a result, data-driven approaches based on deep neural networks have been widely explored as alternatives for long-horizon dynamical prediction, including Recurrent Neural Networks (RNNs) [ 7 ] and Long Short-Term Memory (LSTM) networks [ 15 ] . However, their effectiveness in long-horizon forecasting remains limited due to vanishing gradients and cumulative prediction errors, due to their autoregressive structures that sequentially generate future states by conditioning on previously predicted outputs. To alleviate these limitations, subsequent research has explored probabilistic and generative modeling approaches that aim to capture the full predictive distribution rather than producing a single deterministic trajectory. Among these, Generative Adversarial Networks (GANs) [ 12 ] were introduced to enhance sample diversity and model uncertainty in dynamical prediction. Despite their ability to generate diverse samples, however, GAN-based approaches suffer from well-known challenges, including training instability, mode collapse, and limited reproducibility, which hinder their reliability in long-term dynamical forecasting. Diffusion models [ 14 , 28 ] provide a principled alternative to autoregressive and adversarial approaches by learning the full conditional distribution through iterative noise injection and denoising. Their probabilistic nature enables uncertainty estimation, while their connections to ordinary differential equations (ODEs), partial differential equations (PDEs), and stochastic differential equations (SDEs) make them well-suited for physics-informed modeling. Owing to these advantages, diffusion models have become a dominant paradigm in high-dimensional generative modeling, particularly in static image synthesis. Until recently, however, their application to time-dependent and continuously evolving dynamical systems remained limited. Recent studies have begun to extend diffusion models to long-horizon spatiotemporal forecasting, with TimeDiff [ 27 ] , MCVD [ 30 ] , and DYffusion [ 25 ] serving as representative examples. These works demonstrate that diffusion-based frameworks can successfully move beyond static generation and be adapted to dynamic prediction tasks involving long time scales. Despite this progress, several fundamental challenges remain unresolved when applying diffusion models to continuous-time dynamical systems. Existing approaches typically rely on Gaussian diffusion processes and optimize training objectives based primarily on mean squared error, which inadequately capture the intrinsic physical and dynamical structure governed by underlying PDEs. As a consequence, these models may exhibit degraded distributional fidelity, limited point-wise tracking accuracy, and increased sensitivity to noise in long-horizon predictions. Furthermore, Gaussian-based diffusion formulations are not immune to cumulative error effects, particularly when deployed in iterative forecasting settings. Motivated by these limitations, we propose a novel framework named PDYffusion , a PDE-based dynamics-informed diffusion model for long-horizon spatiotemporal forecasting. The proposed framework decomposes the prediction process into two complementary stages: interpolation and forecasting. Specifically, we introduce a PDE-regularized interpolator that constructs physically consistent intermediate states by incorporating differential operators into the interpolation process. On top of this, we design an uncertainty-aware forecaster based on the Unscented Kalman Filter (UKF), which propagates predictions while explicitly modeling uncertainty and mitigating cumulative error. By coupling physics-aware interpolation with uncertainty-aware forecasting, the proposed framework enables stable and coherent long-horizon prediction. In a nutshell, the main contributions of this work are summarized as follows. • We introduce a PDE-regularized interpolation mechanism that enforces physically consistent intermediate trajectories, improving distributional fidelity and temporal coherence beyond standard MSE-based objectives. • We develop an uncertainty-aware forecasting scheme based on the Unscented Kalman Filter, which effectively reduces cumulative error and stabilizes long-term predictions. • We provide theoretical analysis for both components, establishing their convergence properties and supporting the proposed design. The remainder of this paper is organized as follows. In Section 2 , we explain the prerequisite background for our research. In Section 3 , we develop the proposed PDYffusion model using the PDE-based sampling and UKF-based filtering method, along with its mathematical analysis provided in Section 4 . In Section 5 , we present quantitative and qualitative experimental results on long-horizon dynamical datasets, demonstrating the effectiveness of our model. At last, Section 6 concludes the paper. 2 Related Work 2.1 Time-series Prediction using Deep Generative Model Deep generative models, including Variational Autoencoders (VAEs) [ 20 ] , Generative Adversarial Networks (GANs) [ 12 ] , Normalizing Flows [ 24 ] , and diffusion-based models [ 14 , 28 ] , aim to learn the underlying probability distribution of data in order to generate new samples or perform related tasks such as interpolation, restoration, and conditional generation. Among these families, diffusion models have recently emerged as a dominant paradigm due to their training stability, strong mode coverage, and scalability. By learning a gradual noise injection process and its corresponding denoising dynamics, diffusion models avoid adversarial optimization and exhibit favorable quality–diversity trade-offs across a wide range of domains, including images, audio, video, and scientific data. In most existing formulations, diffusion models are constructed using Gaussian forward processes and are trained by minimizing mean-squared error (MSE) objectives derived from denoising score matching. Several recent works have extended Gaussian-based diffusion models with the backward reconstruction principle of Denoising Diffusion Probabilistic Models (DDPM) [ 14 ] for spatiotemporal and dynamical settings. TimeDiff [ 27 ] and MCVD [ 30 ] adapt the standard diffusion framework to long-horizon prediction by modeling temporal evolution through conditional denoising and masked reconstruction strategies. These approaches demonstrate that diffusion models can be applied beyond static data generation and can capture complex temporal dependencies over extended time scales. However, despite their empirical successes, these methods largely inherit the core assumptions of Gaussian diffusion models, including MSE-based training objectives and Markovian dynamics. As a consequence, existing diffusion-based dynamical models often struggle to capture the underlying physical structure of continuous-time systems governed by partial differential equations. In particular, Gaussian noise assumptions and purely data-driven objectives limit their ability to enforce physical consistency, which leads to degraded long-horizon stability, distributional drift, and sensitivity to noise accumulation. These limitations motivate the incorporation of physics-informed structure and uncertainty-aware correction