Particle-Guided Diffusion Models for Partial Differential Equations

We introduce a guided stochastic sampling method that augments sampling from diffusion models with physics-based guidance derived from partial differential equation (PDE) residuals and observational constraints, ensuring generated samples remain physically admissible. We embed this sampling procedure within a new Sequential Monte Carlo (SMC) framework, yielding a scalable generative PDE solver. Across multiple benchmark PDE systems as well as multiphysics and interacting PDE systems, our method produces solution fields with lower numerical error than existing state-of-the-art generative methods.

💡 Research Summary

The paper introduces a novel generative PDE solver that combines pretrained diffusion models with physics‑based guidance derived from PDE residuals and sparse observations, and embeds this guided sampling within a Sequential Monte Carlo (SMC) framework. The authors first train a diffusion model on joint distributions of PDE coefficients (a) and solution fields (u), treating each pair (a, u) as a single data point. During inference, they define a log‑likelihood that penalizes deviations from observed coefficients and solutions as well as the squared norm of the PDE residual f(c, τ, x). This likelihood is incorporated as a guidance term in the reverse‑time stochastic differential equation (SDE) or ordinary differential equation (ODE) that drives the diffusion sampling, effectively tilting the unconditional diffusion prior toward the posterior p(x | y) ∝ p(y | x)p(x).

To improve the approximation of the posterior, the authors adopt an SMC algorithm that propagates a population of N particles through discrete time steps. At each step, particles are proposed using a kernel Mₖ and re‑weighted by a potential function Gₖ that reflects the guidance information. The key methodological contribution is the introduction of a second‑order stochastic proposal augmented with PDE guidance (SOSaG). SOSaG leverages second‑order information to generate more informed proposals that respect the physics‑based guidance, leading to lower variance weights and higher effective sample sizes. However, the authors show that SOSaG is incompatible with the strict unbiasedness requirements of classical SMC; consequently, they propose a hybrid “vanilla guidance + SMC” scheme that deliberately sacrifices statistical consistency for empirical performance gains.

Extensive experiments are conducted on six benchmark problems, including static Darcy flow, time‑dependent Navier‑Stokes, and two‑ and three‑species interacting PDE systems. The proposed methods (SOSaG‑Prop and a generalized EM‑based proposal) consistently achieve lower relative errors than state‑of‑the‑art guided diffusion approaches such as DiffusionPDE and G‑DiffPDE. For example, SOSaG‑Prop reduces the average error to around 0.8 % compared with 3–5 % for the baselines, and it remains robust when observations are extremely sparse. The paper also provides qualitative visualizations that demonstrate superior reconstruction of coefficient fields and solution fields.

The contributions can be summarized as follows: (1) a physics‑informed likelihood that jointly encodes PDE residuals and observation masks, (2) an SMC‑based guided sampling framework tailored to high‑dimensional diffusion models, (3) the SOSaG proposal that integrates second‑order stochastic dynamics with PDE guidance, (4) a pragmatic relaxation of SMC’s unbiasedness to achieve better empirical results, and (5) comprehensive empirical validation across a variety of PDE settings showing significant error reductions.

While the work advances the integration of generative diffusion models with physical constraints, several open issues remain. The trade‑off between statistical consistency and empirical performance is not theoretically quantified, raising questions about convergence guarantees as the number of particles grows. The sensitivity of the method to the weighting hyperparameters (β, γ, ω) and to the choice of noise schedule is only briefly explored. Moreover, scalability to large‑scale three‑dimensional problems and real‑time data assimilation scenarios warrants further investigation. Future research directions include rigorous analysis of the hybrid SMC bias, adaptive tuning of guidance weights, and extension to multi‑physics coupling with heterogeneous discretizations.