Abstract:We introduce a general framework for solving partial differential equations (PDEs) using generative diffusion models. In particular, we focus on the scenarios where we do not have the full knowledge of the scene necessary to apply classical solvers. Most existing forward or inverse PDE approaches perform poorly when the observations on the data or the underlying coefficients are incomplete, which is a common assumption for real-world measurements. In this work, we propose DiffusionPDE that can simultaneously fill in the missing information and solve a PDE by modeling the joint distribution of the solution and coefficient spaces. We show that the learned generative priors lead to a versatile framework for accurately solving a wide range of PDEs under partial observation, significantly outperforming the state-of-the-art methods for both forward and inverse directions.

What problem does this paper attempt to address?

This paper attempts to solve the problem of solving partial differential equations (PDEs) under partial observation conditions. Specifically, the paper focuses on how to use generative diffusion models to simultaneously fill in missing information and solve PDEs in the absence of complete scene knowledge. Existing forward or reverse PDE methods perform poorly when the observations of data or underlying coefficients are incomplete, and this situation is very common in actual measurements. The DiffusionPDE framework proposed in this paper can effectively manage uncertainty and simultaneously reconstruct these two spaces by modeling the joint distribution of the solution space and the coefficient space, thereby accurately solving various PDEs under partial observation conditions, significantly outperforming existing methods. ### Main problems 1. **Solving PDEs under partial observation**: - **Forward problem**: Predict the final state of the system given partial initial conditions or coefficients. - **Reverse problem**: Infer the initial conditions or coefficients of the system given partial final states. 2. **Limitations of existing methods**: - **Classical methods**: Rely on complete observation data and have poor ability to handle partial observation data. - **Existing learning methods**: Although fast and able to handle a class of PDEs, they also require complete observation data and are difficult to cope with sparse observation situations in reality. ### Solutions - **DiffusionPDE framework**: Use generative diffusion models to model the joint distribution of the solution space and the coefficient space. Through the iterative denoising process, combine sparse observations and PDE constraints to restore the complete initial conditions and final states. - **Algorithm flow**: - **Pre - training**: Pre - train the diffusion generative model on a series of PDEs to learn the joint distribution. - **Inference process**: At the time of inference, start from Gaussian noise, gradually denoise, and add the guidance of sparse observations and PDE constraints at each step, and finally restore the complete solution and coefficients. ### Experimental verification - **Experimental setup**: Tested multiple static and dynamic PDEs, including Darcy Flow, Poisson equation, Helmholtz equation, Burger’s equation, and Navier - Stokes equation. - **Performance comparison**: Compared with existing methods (such as PINO, DeepONet, PINNs, and FNO), DiffusionPDE shows significant advantages under partial observation conditions, especially in reverse problems. ### Conclusion DiffusionPDE successfully solves the problem of solving PDEs under partial observation conditions through generative diffusion models and joint distribution learning, demonstrating its potential in practical applications, especially the ability to accurately restore the system state under sparse observation data.

DiffusionPDE: Generative PDE-Solving Under Partial Observation

Generative Diffusion Learning for Parametric Partial Differential Equations

Diffusion models as probabilistic neural operators for recovering unobserved states of dynamical systems

Diffusion with Forward Models: Solving Stochastic Inverse Problems Without Direct Supervision

ODE-DPS: ODE-based Diffusion Posterior Sampling for Inverse Problems in Partial Differential Equation

Generative Learning of the Solution of Parametric Partial Differential Equations Using Guided Diffusion Models and Virtual Observations

Diffusion model based data generation for partial differential equations

A Score-based Generative Solver for PDE-constrained Inverse Problems with Complex Priors

On conditional diffusion models for PDE simulations

A Flexible Diffusion Model

Diffusion Posterior Sampling for General Noisy Inverse Problems

DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics

Learning to Diffuse: A New Perspective to Design PDEs for Visual Analysis

Learning Diffusion Priors from Observations by Expectation Maximization

Generative downscaling of PDE solvers with physics-guided diffusion models

Improvement of DPM Solver Based on Pade Approximation

Text2PDE: Latent Diffusion Models for Accessible Physics Simulation

A Geometric Perspective on Diffusion Models

Diffusion Probabilistic Fields

Learning Partial Differential Equations for Computer Vision ∗