---
title: 'PDYffusion: PDE-informed Diffusion Forecasting'
url: https://www.emergentmind.com/topics/pdyffusion
type: topic
---

# PDYffusion: PDE-informed Diffusion Forecasting

PDYffusion, introduced in “PDE-regularized Dynamics-informed Diffusion with Uncertainty-aware Filtering for Long-Horizon Dynamics,” is a dynamics-informed diffusion framework for long-horizon spatiotemporal forecasting that couples a PDE-regularized interpolator with a UKF-based forecaster [2604.09058]. It is designed for high-dimensional dynamical systems whose trajectories are naturally described by PDEs, and targets the regime in which cumulative prediction error, noise amplification, and loss of physical consistency make long-term prediction difficult. The method adopts the interpolation-plus-forecasting structure of DYffusion, but replaces pure MSE interpolation with PDE-constrained regularization and replaces heuristic uncertainty handling with an Unscented Kalman Filter trained through a likelihood-based objective [2306.01984].

## 1. Problem formulation and design goals

PDYffusion considers forecasting trajectories of spatiotemporal fields over long horizons. A trajectory is written as
\[
x_{t:t+h} = (x_t, x_{t+1}, \ldots, x_{t+h}),
\]
where \(x_t\) is the observed state and the objective is to predict \(x_{t+1:t+h}\) from \(x_t\) [2604.09058]. The experimental settings include Sea Surface Temperature (SST), Navier–Stokes, Spring-mesh, and Wave, all of which are systems naturally described by PDEs and evaluated in long-horizon regimes.

The method is motivated by three recurrent failure modes of long-term forecasting. First, autoregressive predictors accumulate error because each rollout step conditions on previous predictions. Second, stochasticity and model noise are repeatedly propagated, yielding noise amplification. Third, objectives based only on pointwise reconstruction do not enforce PDE constraints, so trajectories can drift away from physically plausible manifolds [2604.09058].

The paper positions these limitations as only partially addressed by prior diffusion-based forecasters. In particular, DYffusion uses interpolation and prediction trained with pure MSE on states and does not explicitly model uncertainty propagation across time beyond sampling noise [2306.01984]. PDYffusion is designed to address these limitations through two linked mechanisms: a PDE-regularized interpolator that enforces physically consistent intermediate states, and a UKF-based forecaster that explicitly propagates mean and covariance during iterative prediction [2604.09058].

## 2. Framework architecture and dynamics-informed diffusion view

PDYffusion consists of two key components. The first is a PDE-regularized interpolator \(I_\phi\), which reconstructs intermediate states between two time points:
\[
\hat{x}_{t+i} = I_\phi(x_t, x_{t+h}, i),
\]
for \(i \in \{1,\dots,h-1\}\). The second is a UKF-based forecaster \(F_\theta\), which predicts the terminal state \(x_{t+h}\) while maintaining an uncertainty estimate \(\mathcal{N}(\hat{x}_{t+h}, P_{t+h})\) [2604.09058].

Conceptually, the framework inherits DYffusion’s reinterpretation of diffusion as interpolation plus forecasting rather than Gaussian noising plus denoising [2306.01984]. In PDYffusion, the interpolation network plays the role of the forward process, the forecaster plays the role of the reverse process, and the sampling procedure is cold and deterministic in the sense that there is no explicit Gaussian noise schedule or learned \(\epsilon_\theta(x_t,t)\) term [2604.09058].

This diffusion view is therefore not DDPM-like. The method has no discrete-time Gaussian forward kernel \(q(x_t\mid x_{t-1})\), no variance schedule, and no reverse denoiser trained to predict additive noise. Instead, the interpolation stage is regularized by the spatial differential operator
\[
\bigl(E-l^{-2}\Delta\bigr)^{\alpha/2},
\]
and the forecasting stage is regularized by a UKF-based likelihood that couples mean accuracy with covariance calibration [2604.09058].

A plausible implication is that the term “diffusion” in PDYffusion is primarily structural and algorithmic rather than tied to classical score-based diffusion. The method remains diffusion-inspired through the time-indexed interpolation-and-refinement process, but its inductive bias is explicitly adapted to spatiotemporal dynamics.

## 3. PDE-regularized interpolator

The mathematical setting assumes that data arise from PDE-governed dynamics of the form
\[
\partial_t u(t,x) = \mathcal{F}\big(u(t,x), \nabla u(t,x), \nabla^2 u(t,x), x, t\big),
\]
over a spatial domain \(\Omega \subset \mathbb{R}^d\) and a time interval \([t,t+h]\) [2604.09058]. Rather than imposing a task-specific PDE, PDYffusion uses a generic elliptic operator inspired by SPDE/PDE-based sampling with Matérn covariance.

Let \(\Delta\) denote the spatial Laplacian, \(l\) a correlation length, and \(\alpha\) a smoothness parameter. The method defines
\[
M := \bigl(E - l^{-2}\Delta\bigr)^{\alpha/4}, \qquad
M^2 = \bigl(E - l^{-2}\Delta\bigr)^{\alpha/2}.
\]
Interpolation is associated with minimizing
\[
J(u) = \frac{1}{2} \int_t^{t+h} \int_{\Omega}
\left| \bigl(E - l^{-2}\Delta\bigr)^{\alpha/4} u(x,t) \right|^2 \, dx\,dt
\]
subject to boundary conditions
\[
u(x,t)=x_t,\qquad u(x,t+h)=x_{t+h}.
\]
Using self-adjointness of \(\Delta\), the Euler–Lagrange condition yields
\[
\bigl(E - l^{-2}\Delta\bigr)^{\alpha/2} u(x,t)=0,
\]
so the ideal interpolating field lies in the null space of the elliptic operator [2604.09058].

The neural interpolator is trained to approximate such a field through the loss
\[
\min_{\phi}\;
\mathbb{E}\Big[
\|I_{\phi}(x_t,x_{t+h},i)-x_{t+i}\|^{2}
+ \lambda \|(E-l^{-2}\Delta)^{\alpha/2}I_{\phi}\|^{2}
\Big].
\]
This combines reconstruction with PDE regularization. The paper states that the interpolator therefore learns intermediate states that are not only pointwise accurate but also consistent with an elliptic PDE in space [2604.09058].

Theoretical analysis is given for this regularization. Using the SPDE–Matérn equivalence, the paper argues that the Matérn kernel induced by the operator is positive semi-definite by Bochner’s theorem, which enables MMD analysis. It further states that PDE regularization reduces error covariance in Loewner order relative to a purely MSE-trained interpolator and derives
\[
\mathbb{E}\big[\mathrm{MMD}_k^2(P_\phi, P_u)\big]
\le
\mathbb{E}\big[\mathrm{MMD}_k^2(P_{\mathrm{base}}, P_u)\big],
\]
indicating that the PDE-regularized interpolator distribution is closer in RKHS sense to the true dynamics distribution than the baseline interpolator [2604.09058].

## 4. UKF-based forecaster and uncertainty-aware filtering

The forecaster treats long-horizon prediction as nonlinear state estimation with explicit uncertainty. The dynamics model is written as
\[
x_k = f(x_{k-1}) + w_k,
\]
where \(f\) is implicitly modeled by \(F_\theta\) and \(w_k \sim \mathcal{N}(0,Q)\), while observations satisfy
\[
z_k = h(x_k) + v_k,
\]
with \(v_k \sim \mathcal{N}(0,R)\) [2604.09058].

The UKF propagates sigma points through the learned dynamics. Given a state estimate \(\hat{x}_k\) and covariance \(P_k\), sigma points are generated as
\[
\mathcal{X}_k^{(0)} = \hat{x}_k,
\]
\[
\mathcal{X}_k^{(i)} = \hat{x}_k + \bigl[\sqrt{(n+\lambda)P_k}\bigr]_i,\quad i=1,\dots,n,
\]
\[
\mathcal{X}_k^{(i)} = \hat{x}_k - \bigl[\sqrt{(n+\lambda)P_k}\bigr]_{i-n},\quad i=n+1,\dots,2n,
\]
with standard UKF weights determined by \(\lambda = \alpha^2(n+\kappa)-n\). These sigma points are propagated through \(F_\theta\), producing a predicted mean and covariance, then mapped through the measurement function \(h\), and finally updated with the Kalman gain
\[
G = P_{xz} S_{k+1}^{-1}
\]
to obtain posterior mean and covariance [2604.09058].

The terminal forecast is modeled as
\[
x_{t+h} \sim \mathcal{N}(\hat{x}_{t+h}, P_{t+h}),
\]
and training uses the negative log-likelihood
\[
L_{\mathrm{UKF}}(\theta,\phi)
=
-\log p(x_{t+h}\mid \hat{x}_{t+h}, P_{t+h})
=
\frac{d}{2}\log(2\pi)
+
\frac{1}{2}\log|P_{t+h}|
+
\frac{1}{2} e_{t+h}^{\top} P_{t+h}^{-1} e_{t+h},
\]
where \(e_{t+h}=x_{t+h}-\hat{x}_{t+h}\) [2604.09058].

The full forecasting loss is
\[
\min_{\theta,\phi}
\Big(
\|F_{\theta}(I_{\phi}, i_n) - x_{t+h}\|_2^2
+
L_{\mathrm{UKF}}(\theta,\phi)
\Big).
\]
The MSE term enforces accurate mean prediction, the Mahalanobis term penalizes errors relative to predicted uncertainty, and the log-determinant term penalizes overly large or small covariance [2604.09058]. The paper further gives a convergence result stating that the objective tends to zero if and only if mean prediction converges, sigma-point covariance collapses appropriately, and \(P_{t+h}\to Q\) with \(|Q|=(2\pi)^{-d}\).

## 5. Training procedure, inference, and empirical behavior

Training proceeds in two stages. First, the interpolator \(I_\phi\) is trained with the PDE-regularized reconstruction loss on triplets \((x_t, x_{t+i}, x_{t+h})\). Second, the interpolator is frozen, a diffusion step index \(n\) is sampled uniformly, intermediate inputs related to \(i_n\) are generated, UKF prediction and update are run around \(F_\theta(I_\phi, i_n)\), and the forecaster is optimized with the combined MSE-plus-UKF objective [2604.09058].

Inference mirrors DYffusion’s cold sampling structure but augments it with uncertainty-aware state estimation [2306.01984]. Starting from the observed state \(x_t\) and an initialized covariance \(P_t\), the method iteratively selects a time index \(i_n\), computes
\[
\hat{x}_{t+i_n} = I_\phi(x_t, \hat{x}_{t+h}^{(n)}, i_n),
\]
then updates the terminal forecast and covariance through the forecaster and UKF to obtain \((\hat{x}_{t+h}^{(n+1)}, P_{t+h}^{(n+1)})\) [2604.09058].

Evaluation uses MSE, CRPS, and SSR. MSE measures pointwise accuracy; CRPS measures the predictive distribution; SSR is interpreted with values near \(1\) as ideal, \(<1\) as underdispersive, and \(>1\) as overdispersive [2604.09058].

The principal quantitative results summarized in the paper are as follows. On Navier–Stokes, PDYffusion achieves best CRPS \(0.059\) and best MSE \(0.017\), with SSR \(0.731\). On Spring-mesh, it achieves best CRPS \(0.0092\), best MSE \(4.01\mathrm{e}{-4}\), and SSR \(1.204\). On Wave, it achieves best CRPS \(1.82\mathrm{e}{-3}\), best MSE \(1.57\mathrm{e}{-5}\), and best SSR \(0.913\), which is the closest to \(1\) among all methods. On SST, MCVD performs slightly better in CRPS and MSE, while PDYffusion remains competitive with CRPS \(0.226\) and MSE \(0.166\) [2604.09058].

The paper also reports two informative ablations. For boundary conditions in the PDE operator, periodic boundary conditions perform best on SST and Wave, while Dirichlet performs best on Navier–Stokes and Spring-mesh; Neumann is typically intermediate. For regularization strength on SST, \(\lambda=0.2\) gives best CRPS \(0.226\), best MSE \(0.166\), and SSR \(0.683\), whereas stronger regularization such as \(\lambda=0.5\) or \(1.0\) degrades performance [2604.09058].

Qualitatively, the Wave experiments show that both DYffusion and PDYffusion are close to ground truth at early times, but at long times DYffusion develops irregular oscillations and noisy deviations, whereas PDYffusion maintains smooth wave shapes with reduced spurious oscillation [2604.09058].

## 6. Relation to adjacent methods, limitations, and common sources of confusion

PDYffusion is most directly related to DYffusion. Both methods replace classical noise-driven diffusion with an interpolation-plus-forecasting process aligned with physical time, and both use cold sampling rather than a DDPM-style forward corruption schedule [2306.01984]. The methodological distinction is that PDYffusion adds a PDE-constrained variational loss to the interpolator and a UKF-based uncertainty model to the forecaster [2604.09058].

The name can also be confused with “Pyramid Diffusion,” or Pyramid Discrete Diffusion, which is a multi-scale discrete diffusion framework for generating large outdoor 3D semantic scenes with a coarse-to-fine pyramid of voxelized semantic grids and a scene subdivision mechanism [2311.12085]. That method addresses 3D scene generation under memory and data constraints, not long-horizon spatiotemporal forecasting. This suggests that the similarity in naming is nominal rather than methodological.

The strengths identified for PDYffusion are physics-aware interpolation, uncertainty-aware forecasting, improved long-horizon performance, and clear interpretability of the parameters \(l\) and \(\alpha\) as correlation length and smoothness [2604.09058]. The paper simultaneously notes several limitations. The PDE regularizer uses a generic Matérn-like elliptic operator with Laplacian and may not precisely match dynamics involving advection, nonlinearity, or non-elliptic operators. PDE operator evaluation and UKF covariance operations introduce computational overhead. The implementation relies on discretized Laplacians and boundary conditions, making extension to irregular meshes or very high spatial dimensions nontrivial. Finally, the interpolator is deterministic, and the only explicit probabilistic component is the UKF Gaussian [2604.09058].

The extensions proposed in the paper include moving from deterministic PDE systems with observational noise to intrinsically stochastic systems such as SPDEs and SDEs, learning unknown differential operators from data rather than fixing \((E-l^{-2}\Delta)^{\alpha/2}\), scaling to 3D flows and complex geometries, and combining the framework with richer probabilistic models such as VAEs, normalizing flows, or alternative Bayesian filtering schemes when Gaussian assumptions are inadequate [2604.09058].

Source: https://www.emergentmind.com/topics/pdyffusion