A Score-based Diffusion Model Approach for Adaptive Learning of Stochastic Partial Differential Equation Solutions
Toan Huynh, Ruth Lopez Fajardo, Guannan Zhang, Lili Ju, Feng Bao · 2025
Read A Score-based Diffusion Model Approach for Adaptive Learning of Stochastic Partial Differential Equation Solutions by Toan Huynh, Ruth Lopez Fajardo, Guannan Zhang, Lili Ju, Feng Bao (2025) 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-2508-06834.txt.
Introduction
Toan Huynh and co-authors propose, in a 2025 paper, a learning method for equations that describe physical systems buffeted by chance. Such stochastic partial differential equations model everything from fluids to materials under uncertainty, but their computed solutions drift when the physics is partly unknown or the setting shifts. The authors aim to correct the solution as fresh data arrives.
Their framework bakes the physics into the score function of a diffusion model trained on simulations. Observations then enter through a likelihood step inside a reverse-time equation, so each new batch of data refines the answer. A training-free shortcut, the ensemble score filter, keeps the cost sane in high dimensions and suits real-time use. On benchmark equations with sparse, noisy data, the method stayed accurate and robust.
Catalog record
| CREATOR | Toan Huynh, Ruth Lopez Fajardo, Guannan Zhang, Lili Ju, Feng Bao |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-06834
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Score-based Diffusion Model Approach for Adaptive Learning of Stochastic Partial Differential Equation Solutions Toan Huynh, Ruth Lopez Fajardo, Guannan Zhang, Lili Ju, Feng Bao Abstract We propose a novel framework for adaptively learning the time-evolving solutions of stochastic partial differential equations (SPDEs) using score-based diffusion models within a recursive Bayesian inference setting. SPDEs play a central role in modeling complex physical systems under uncertainty, but their numerical solutions often suffer from model errors and reduced accuracy due to incomplete physical knowledge and environmental variability. To address these challenges, we encode the governing physics into the score function of a diffusion model using simulation data and incorporate observational information via a likelihood-based correction in a reverse-time stochastic differential equation. This enables adaptive learning through iterative refinement of the solution as new data becomes available. To improve computational efficiency in high-dimensional settings, we introduce the ensemble score filter, a training-free approximation of the score function designed for real-time inference. Numerical experiments on benchmark SPDEs demonstrate the accuracy and robustness of the proposed method under sparse and noisy observations. 1 Introduction In this paper, we introduce a score-based diffusion model approach for adaptively learning the time-evolving solutions of stochastic partial differential equations (SPDEs) through recursive Bayesian inference. Partial differential equations (PDEs) are fundamental tools for modeling the dynamic behavior of complex physical systems. While they have been widely successful in scientific and engineering applications, many practical scenarios involve inherent uncertainties due to limited physical knowledge and environmental variability. For example, in climate and meteorological modeling, uncertainties in initial conditions, boundary data, and subgrid-scale physical processes can significantly affect the accuracy of predictions governed by PDEs such as the Navier–Stokes or advection–diffusion equations. Similarly, in porous media flow problems, spatial heterogeneity and limited characterization of subsurface properties — such as permeability or porosity — introduce substantial uncertainty into models governed by Darcy’s law and related PDEs, making accurate prediction particularly challenging. To capture these uncertainty effects and support reliable predictive analysis, it is essential to incorporate SPDEs into mathematical modeling framework. The numerical solution of SPDEs has thus become a central focus of the uncertainty quantification (UQ) community, where significant efforts have been dedicated to developing efficient solvers that can accurately characterize and propagate uncertainty in high-dimensional, nonlinear dynamical systems (see, e.g., [ 1 , 2 , 13 , 21 , 36 , 42 , 53 ] and the reference therein). Despite advances in SPDE solvers capable of quantifying uncertainty, significant challenges remain. In particular, model errors arising from limited physical knowledge and uncertainties due to unknown environmental variability can accumulate over time, leading to a progressive degradation in solution accuracy — especially when the underlying physical laws are incomplete or only approximately known [ 9 , 15 , 20 , 30 , 31 , 35 , 41 , 45 ] . A strategy to address these issues is to leverage measurement data to “learn” an estimated solution that reflects the underlying physical reality. However, fully data-driven approaches are often impractical due to the fact that real-world observations are typically sparse, indirect, and contaminated with noise. As a result, there is a critical need for adaptive learning frameworks that can effectively integrate partial and noisy measurement data into existing physical models . Such frameworks are essential for correcting model discrepancies, reducing uncertainty, and ultimately improving the fidelity of SPDE solvers. The standard framework for integrating data and models is known as “data assimilation”, which systematically combines observational data with numerical model simulations to estimate the evolving state of a dynamical system. The mathematical formulation of this problem is referred to as “optimal filtering”, which seeks to represent the “optimal” estimate of the target model state in the form the conditional expectation – conditioning on the observational information. The state-of-the-art approach for solving the optimal filtering problem is the recursive Bayesian filter, including the Kalman type filters [ 14 , 20 , 33 ] and the particle filters [ 3 , 5 , 22 , 34 , 40 , 43 ] . In addition to the Bayesian approach, there exists a class of methods based on stochastic partial differential equations that directly approximate the conditional distribution of the target state [ 4 , 6 , 7 , 8 , 56 , 57 ] . In this work, we develop a novel framework that applies score-based diffusion models to the adaptive learning of SPDE solutions. Following the recursive Bayesian framework, our method used numerical SPDE solvers to encode first-principle and model-based dynamics for predictive simulation. When observational data becomes available, we apply Bayesian inference to integrate this data into the simulated model, enabling calibration of the model-based estimates. The adaptive learning in this context refers to the iterative refinement of the solution: as new data is assimilated, the model dynamically adjusts its estimates, progressively capturing hidden physics and reducing epistemic uncertainty. This leads to improved predictive accuracy and robustness over time, even in the presence of incomplete or imperfect knowledge of the underlying system. The central contribution of this work is the introduction of a score-based diffusion model technique [ 26 , 38 , 51 ] that enables this adaptive learning process through data assimilation. While conventional optimal filtering methods have achieved considerable success, these methods face fundamental limitations. Kalman-type filters are efficient in high-dimensional settings but rely on linear or weakly nonlinear assumptions [ 29 , 20 ] . Particle filters, on the other hand, are well-suited for nonlinear systems but suffer from the “curse of dimensionality” issue in high-dimensional spaces due to its sequential Monte Carlo mechanism [ 28 , 49 ] . As a result, no existing method can effectively address the joint challenges of high dimensionality (e.g., arising from spatial discretization in SPDEs) and strong nonlinearity — both of which are intrinsic to adaptive learning tasks for SPDEs. Our framework is designed to overcome this gap by taking the advantage of the flexibility and scalability of score-based generative models [ 50 ] . In a score-based diffusion model, a forward stochastic differential equation (SDE) is used to progressively transform an arbitrary data distribution into a standard Gaussian distribution. During this process, the information from the original data is encoded in the so-called “score function”, defined as the gradient of the log-probability density of the forward SDE. Then, a reverse-time SDE, driven by the score function, is solved to effectively transform samples drawn from the standard Gaussian distribution back into samples that follow the original data distribution. When applying score-based diffusion models to the adaptive learning of SPDE solutions, we encode the underlying physics described by the SPDE into the score function using simulation samples. When observational data becomes available, a likelihood score is introduced to incorporate data information into the reverse-time SDE, enabling the model to iteratively adjust its estimates and improve solution accuracy [ 10 ] . To enhance the computational efficiency of score-based diffusion models in high-dimensional settings and enable real-time calibration of estimated solutions, we propose an ensemble-based approximation of the score function—rather than relying on neural network-based deep learning methods [ 11 ] . We refer to this approach as the ensemble score filter. We validate the proposed method through a series of numerical experiments involving prototypical