---
title: Physics-Informed Diffusion Models
url: https://www.emergentmind.com/topics/physics-informed-diffusion
type: topic
---

# Physics-Informed Diffusion Models

to=arxiv_search.search  天天种彩票json
{"query":"all:\"Physics-Informed Diffusion Models\" OR all:\"Pi-fusion\" OR all:\"RMDM\" OR all:\"MRF-DiPh\" OR all:\"Towards Physics-informed Diffusion for Anomaly Detection in Trajectories\"","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}Japgollyassistant to=arxiv_search.search
to=arxiv_search.search აციjson
{"query":"all:\"physics-informed diffusion\"","max_results":20,"sort_by":"submittedDate","sort_order":"descending"}
Physics-informed diffusion denotes a class of methods that couple diffusion-based generative modeling or denoising with explicit physical structure such as governing PDEs, constitutive laws, measurement operators, kinematic constraints, or biophysical forward models. In one explicit formulation, a denoising diffusion model is trained not only with the usual diffusion objective but also with a first-principle-based term that enforces generated samples to fulfill physical constraints [2403.14404]. In other formulations, diffusion sampling is guided by proximal physics steps, architecture-level embeddings of physical simulators, or trajectory-level rewards derived from residual operators [2506.23311]. The expression is also used in a broader sense for PINN-based treatment of physical diffusion equations and inverse diffusivity recovery, where “diffusion” refers to the underlying transport process rather than to a DDPM-like generative model [2403.03970]. Across radio propagation, MRI, fluid dynamics, trajectory modeling, multivariate time series, net-load synthesis, infrared translation, and relativistic transport, the common objective is to combine high-dimensional distribution learning with explicit physical validity [2501.19160].

## 1. Scope and conceptual variants

Physics-informed diffusion is not a single algorithmic template. In scientific machine learning, one canonical view treats valid samples \(x_0\) as fields that must satisfy governing equations and boundary conditions, written abstractly as \(\boldsymbol{\mathcal{F}}[u(\boldsymbol{\xi})]=\boldsymbol{0}\) in \(\Omega\) and \(\boldsymbol{\mathcal{B}}[u(\boldsymbol{\xi})]=\boldsymbol{0}\) on \(\partial\Omega\), with a discrete residual operator \(\boldsymbol{\mathcal{R}}(x_0)\) that should vanish [2403.14404]. Under this view, diffusion supplies a generative prior over complex solution manifolds, while physics supplies admissibility criteria.

A second major variant uses diffusion as a prior inside an inverse problem. MRF-DiPh formulates multiparametric MRI reconstruction as \(\min_{x\in\mathcal{B}} f(x)+\lambda h(x)\), where \(f(x)=\|y-\mathcal{A}x\|_2^2\) enforces k-space consistency and \(\mathcal{B}\) enforces Bloch-model consistency; the diffusion model realizes the prior term \(h(x)\) inside an HQS/ADMM scheme [2506.23311]. In this setting, “physics-informed” means that reverse diffusion is not free-running, but repeatedly corrected by acquisition and signal-model constraints.

A third variant embeds physics directly in architecture and conditioning. RMDM uses a dual U-Net architecture in which the first U-Net is physics-informed through Helmholtz-equation residuals, boundary conditions, and source constraints, while the second U-Net performs DDPM-style denoising conditioned on the first stage [2501.19160]. A related but distinct construction appears in synthetic net-load generation, where the denoising network explicitly embeds a solar photovoltaic system performance model and PV generation basis profiles rather than imposing a separate residual penalty [2406.01913].

A fourth variant uses physics as a regularizer on temporal or geometric dynamics. Pi-DPM for anomaly detection in trajectories augments the diffusion objective with a kinematic-bicycle-model loss \(\mathcal{L}_{\text{Pi-DPM}}=\gamma_1\mathcal{L}_{\text{VLB}}+\gamma_2\mathcal{L}_{\text{Rec}}+\gamma_3\mathcal{L}_{\text{Phy}}\) [2506.06999]. TPIDM for multivariate time-series anomaly detection similarly adds a weighted physics-informed loss to the diffusion ELBO, with the weight decaying across diffusion time so that low-noise steps carry most of the physical supervision [2508.11528].

This suggests that physics-informed diffusion is best understood as a family of couplings between diffusion priors and explicit operators \( \mathcal{G}, \mathcal{R}, \mathcal{A}, \Pi_{\mathcal{C}} \), rather than as a fixed training rule.

## 2. Mathematical formulations

A central training-time formulation is the physics-informed diffusion model of Bastek et al., where the practical objective is
\[
L_{\text{PIDM}}(\theta)=\mathbb{E}\left[\lambda_t\|x_0-\hat{x}_0(x_t,t)\|^2+\frac{1}{2\bar{\Sigma}_t}\|\boldsymbol{\mathcal{R}}(x_0^*(x_t,t))\|^2\right].
\]
Here the standard denoising term is combined with a residual penalty derived from a Gaussian virtual likelihood over \(\boldsymbol{\mathcal{R}}(x_0)\) [2403.14404]. The residual may encode equality constraints, inequality constraints, or auxiliary optimization objectives, and can be computed by finite differences, FEM assembly, or other differentiable operators.

A guidance-time formulation appears in Pi-fusion. There the conditional score is decomposed into a data score and a physics term, yielding a posterior score
\[
\nabla_{\mathbf{x}_s}\log p_s(\mathbf{x}_s\mid\mathbf{g},t)
=
-\frac{\mathcal{L}_{\mathrm{PDEs}}(\hat{\mathbf{x}}_0,\mathbf{g},t)}{\sigma_{\mathrm{PDE}}^2}
\frac{\partial \mathcal{L}_{\mathrm{PDEs}}(\hat{\mathbf{x}}_0,\mathbf{g},t)}{\partial \mathbf{x}_s}
+
s_\theta,
\]
which is then inserted into the DDIM update through a physics-informed estimate \(\hat{\mathbf{x}}_0^{\mathrm{Pi}}\) [2406.03711]. Physics here enters both as a training loss and as an inference-time corrective score.

Inverse-problem formulations are often expressed through operator splitting. In MRF-DiPh, the diffusion denoiser serves as the prior proximal
\[
\tilde{x}_t = \operatorname{prox}_{\sigma^2 h}(\hat{x}_t),
\]
while the next step solves
\[
\hat{x}_{t-1} = \arg\min_{x\in \mathcal{B}} f(x)+\frac{\mu}{2}\|x-\tilde{x}_t\|_2^2,
\]
followed by dictionary projection and controlled noise reinjection [2506.23311]. The deterministic component of the DDIM update is computed from the physics-consistent state \(z_{t-1}\), which biases sampling toward the intersection of the learned prior manifold and the physics-constrained manifold.

Reward-based formulations replace value approximations with direct optimization of physical rewards. PIRF casts diffusion sampling as an MDP with terminal reward
\[
r(\boldsymbol{x})=-\|\mathcal{R}(\boldsymbol{x})\|_2^2,
\]
and optimizes
\[
\theta^*=\arg\max_{\{p_t(\cdot\mid\cdot;\theta)\}}
\mathbb{E}[r(\boldsymbol{x}_0)].
\]
Instead of DPS-style value estimation, PIRF backpropagates the trajectory-level reward through deterministic diffusion trajectories, then stabilizes the update with layer-wise truncated backpropagation and weight-based regularization [2509.20570].

Time-series and trajectory settings often use softer, domain-specific residuals. TPIDM trains with
\[
\mathcal{L}_{\text{TPIDM}}=\mathcal{L}_{\text{DM}}+\mathcal{L}_{\text{PI}},
\]
where \(\mathcal{L}_{\text{PI}}\) is weighted by a static schedule \(\overline{\lambda}_{\text{PI},t}\) that approaches zero for late, high-noise diffusion steps [2508.11528]. Pi-DPM for trajectory anomaly detection uses residuals of the kinematic bicycle model, penalizing deviations of \(\dot{x},\dot{y},\dot{\psi},\dot{v}\) from \(v\cos\psi, v\sin\psi, v\kappa, a\), respectively [2506.06999].

## 3. Constraint classes and coupling mechanisms

The physical content injected into diffusion models varies substantially across domains, but the coupling mechanisms are recurrent: residual penalties, projections, attention-based conditioning, embedded simulators, and post-denoising correction.

| Domain and paper | Physical structure | Diffusion coupling |
|---|---|---|
| Radio maps [2501.19160] | Helmholtz equation, Dirichlet BC, source constraints | PINN-conditioned dual U-Net with DDPM denoiser |
| Quantitative MRI [2506.23311] | k-space operator \(\mathcal{A}\), Bloch response manifold \(\mathcal{B}\) | DDM prior inside HQS/ADMM with dictionary projection |
| Fluid dynamics [2406.03711] | Incompressible Navier–Stokes residuals | Conditional DDPM/DDIM with physics-guided sampling |
| Trajectories [2506.06999] | Kinematic bicycle model | DDPM with physics regularization in training |
| Multivariate time series [2508.11528] | Lotka–Volterra, Ohm’s law, inverse dynamics, ideal gas law | Weighted physics-informed loss over diffusion steps |
| Net load [2406.01913] | PVSPM and PV generation basis profiles | Physics-informed denoiser embedding |
| Infrared translation [2407.09299] | TeV decomposition from radiative transfer and emissivity modeling | Latent diffusion with physical reconstruction and TeV-space consistency losses |
| Scientific PDE generation [2403.14404] | PDE and boundary residual operator \(\boldsymbol{\mathcal{R}}(x_0)\) | Diffusion loss plus virtual-likelihood residual term |

Additional formulations extend the same pattern. InterDiff for 3D human–object interactions uses diffusion for future-sequence generation and an interaction-correction module that selects contact-based reference frames and applies contact and penetration-aware corrections during denoising [2308.16905]. PIGMENT learns a universal generative prior over diffusion-MRI microstructure maps using a 3D VQ-VAE and latent diffusion conditioned on \(S_0\), then adapts it zero-shot to subject data through a dual-constraint objective combining signal-model consistency with manifold consistency [2606.00156]. RMDM makes the coupling particularly explicit: a physics-informed conditional predictor produces a coarse radio map, and a conditional DDPM denoiser restores finer structure while remaining anchored to the physics-aware representation [2501.19160].

The constraint class need not be a PDE. In MRI, the relevant physics consists of measurement consistency and Bloch dynamics; in trajectories it is kinematics; in infrared translation it is an imaging equation based on emissivity, integrated thermal radiation, and environment reflection; in net-load synthesis it is a photovoltaic performance model embedded in the denoiser [2506.23311].

## 4. Representative domains

Radio propagation provides a prototypical PDE-plus-diffusion example. RMDM models the radio map \(u(x,y)\) with the 2D Helmholtz equation
\[
\nabla^2 u(x,y)+k^2(x,y)u(x,y)=-f(x,y),
\]
discretizes the Laplacian with central differences, and penalizes PDE residual, boundary-condition loss, and source loss before applying a DDPM-style denoiser conditioned through Anchor Condition and RF-SA modules [2501.19160]. The method is designed for both Static RM and Dynamic RM settings on RadioMapSeer, including buildings, transmitters, and vehicles.

MRI reconstruction illustrates a stronger inverse-problem viewpoint. MRF-DiPh operates on time-series magnetization images \(x\in\mathbb{C}^{s\times h\times w}\), with measurement model \(y\approx\mathcal{A}(x)\) and voxelwise Bloch-consistent structure \(x_v=\rho_v\,Bloch(\mathrm{T1}_v,\mathrm{T2}_v)\) [2506.23311]. PIGMENT occupies a different position in the same landscape: instead of learning directly from raw signals, it learns a universal generative prior of human brain microstructure in parameter space, then enforces tensor, kurtosis, or NODDI forward-model consistency during subject-specific adaptation [2606.00156].

Fluid mechanics highlights inference-time guidance. Pi-fusion models the joint field \(\mathbf{x}_0=[\mathbf{u},p]^\top\) and injects Navier–Stokes information both as a PDE loss and as a corrective gradient in DDIM sampling [2406.03711]. By contrast, PIRF removes explicit score/value approximations and optimizes trajectory-level physical rewards directly on PDE benchmarks including Burgers’, Darcy, Helmholtz, Poisson, and Kolmogorov flow [2509.20570].

Trajectory and sequential-data applications broaden the notion of “physics.” Pi-DPM learns normal trajectory manifolds with a diffusion backbone and a kinematic regularizer derived from the kinematic bicycle model, then detects GPS-spoofing anomalies through reconstruction error [2506.06999]. TPIDM uses known ODE or algebraic laws for Predator–Prey, Lenze, EMPS, and Air Compressor datasets, with a static weight schedule that suppresses physics penalties at highly corrupted diffusion steps [2508.11528].

Image-translation and energy-data settings show that the physical operator can be an observation model or a simulator rather than a field equation. PID for RGB-to-infrared translation introduces TeVNet, which decomposes infrared images into emissivity, temperature-related radiation, and environment-related reflection, then imposes both a physical reconstruction loss and a TeV-space consistency loss during latent diffusion training [2407.09299]. The net-load PDM embeds PVSPM and PV generation basis profiles inside the denoiser so that the model separates a data-driven residual component from a physics-based solar component [2406.01913].

## 5. Empirical behavior and technical trade-offs

Across domains, the empirical record is consistently favorable when explicit physics is added to diffusion, although the gains depend on how the coupling is implemented. In radio-map reconstruction, RMDM reports under the Static RM setting NMSE \(=0.0031\), RMSE \(=0.0125\), and SSIM \(=0.978\), versus RadioDiff at NMSE \(=0.0049\), RMSE \(=0.0190\), and SSIM \(=0.9691\); under the Dynamic RM setting it reports NMSE \(=0.0047\), RMSE \(=0.0146\), and SSIM \(=0.968\) [2501.19160]. In its unbalanced sample distribution setting, the same model reaches NMSE \(=0.0022\) and RMSE \(=0.0117\), outperforming Autoencoder, Deep AE, RadioUnet, and RME-GAN.

For accelerated quantitative MRI, MRF-DiPh improves over the purely data-driven conditional diffusion baseline MRF-IDDPM from T1 MAPE \(8.45 \to 6.75\), T2 MAPE \(22.54 \to 18.40\), TSMI NRMSE \(27.26 \to 18.65\), and k-space NRMSE \(36.06 \to 22.82\) [2506.23311]. The paper further reports that the full method is best when Bloch projection is included, indicating that physics beyond measurement consistency is materially important.

Pi-fusion shows that inference-time physics guidance can shift diffusion from a merely data-matching surrogate to a high-accuracy PDE surrogate. On the 2D cylinder benchmark it reports RMSE \(9\times10^{-3}\), nRMSE \(2.23\times10^{-2}\), cRMSE \(1.07\times10^{-2}\), and bRMSE \(1.5\times10^{-3}\), substantially outperforming PINN, NSFnets, PIPN, and ablated variants; on the 3D brain artery benchmark it reports RMSE \(6.51\times10^{-2}\) and nRMSE \(3.18\times10^{-2}\), with inference time \(0.03\) s versus \(90.07\) s for traditional CFD [2406.03711]. For Bastek et al., the central empirical statement is that the residual error is reduced by about two orders of magnitude compared to standard diffusion in the Darcy-flow case study, while topology optimization metrics also improve relative to task-specific baselines [2403.14404].

Sequential-data settings exhibit a similar but more nuanced picture. TPIDM improves F1 on Predator–Prey from \(0.9712\) to \(0.9830\) for the \(\epsilon_\theta\) formulation and from \(0.9957\) to \(1.0\) for the \(x_\theta\) formulation, while remaining competitive or tied on EMPS, Lenze, and Air Compressor [2508.11528]. The net-load PDM reports at least \(20\%\) improvement across all quantitative metrics relative to a well-calibrated baseline diffusion model, with MAE \(0.73\) versus \(1.08\), RMSE \(1.05\) versus \(2.01\), and CRPS \(0.45\) versus \(0.58\) [2406.01913]. PIGMENT extends the same theme to quantitative diffusion MRI at scale, with FA/MD/AD/V1 error reductions on low-field \(0.55\)T data of \(49.8\%\), \(15.5\%\), \(37.3\%\), and \(32.1\%\) relative to OLS, and with strong biological validity under \(10\)-fold accelerated acquisitions [2606.00156].

The trade-offs are equally recurrent. RMDM is described as relatively heavy because it combines dual U-Nets, PINN losses, and diffusion steps [2501.19160]. MRF-DiPh requires diffusion sampling, CG solvers, and dictionary matching, with dictionary matching accounting for about \(40\%\) of runtime and base reconstruction around \(44\) s per \(2\)D slice [2506.23311]. PIRF improves inference efficiency because, unlike guidance-based methods, it incurs zero reward queries and zero model backward passes at inference, but the fine-tuning stage still requires substantial offline compute and careful control of reward hacking [2509.20570].

## 6. Relation to PINNs, ambiguities, and open directions

A persistent ambiguity is that “physics-informed diffusion” can refer either to diffusion models informed by physics or to physics-informed neural solution of physical diffusion equations. The inverse-recovery work on spatially varying diffusivity uses two fully connected networks, one for concentration \(\tilde c(x,y,t)\) and one for diffusivity \(\tilde D(x,y)\), and minimizes a data-plus-physics loss derived from a backward-Euler form of
\[
c_t - c_xD_x - c_yD_y - D(c_{xx}+c_{yy}) = 0,
\]
achieving less than \(6.31\%\) error on all numerical benchmarks [2403.03970]. In expanding-QCD applications, a PINN solves the advection form
\[
\frac{\partial \rho_T}{\partial \tau}
=
-
\frac{\partial(\rho_T v_x)}{\partial x}
-
\frac{\partial(\rho_T v_y)}{\partial y},
\]
with hydrodynamic velocity fields supplied by MUSIC, using curriculum learning and a hybrid Adam/L-BFGS schedule [2601.07120]. For relativistic BDNK diffusion, the SA-PINN-ACTO framework reformulates the system in flux-conservative form, exactly enforces initial and periodic boundary conditions through algebraic transforms, and matches converged Kurganov–Tadmor solutions for smooth profiles while exhibiting larger errors near discontinuities [2602.16117].

This broader usage matters because it clarifies what is novel about generative physics-informed diffusion. PINNs directly approximate a solution field; diffusion models instead learn a distribution or a prior over fields, trajectories, or parameter maps. The literature repeatedly warns against the misconception that a learned generative prior alone is sufficient for physical validity. RMDM explicitly argues that pure DDPM-style radio-map generators may generate unphysical maps under limited data or out-of-distribution geometries, and MRF-DiPh emphasizes that direct regression or purely data-driven diffusion may produce plausible-looking but physically inconsistent maps [2501.19160].

Another recurring misconception is that “physics-informed” always means an explicit PDE residual. The surveyed works show a much wider design space: Bloch-manifold projections and k-space consistency in MRI [2506.23311]; TeV decomposition losses in infrared imaging [2407.09299]; photovoltaic basis embeddings in load synthesis [2406.01913]; kinematic regularization in trajectories [2506.06999]; and reward optimization over terminal PDE residuals in scientific field generation [2509.20570]. This suggests that the operative criterion is not the presence of a PDE term per se, but the insertion of a known physical operator, constraint set, or constitutive law into the diffusion pipeline.

Open directions also recur across papers. Efficiency is a dominant concern: RMDM targets future work on efficiency and scalability, MRF-DiPh points to approximate Bloch projections and alternative diffusion architectures, Pi-fusion notes the cost of multi-step sampling, and PIRF motivates efficient few-step alignment regimes [2501.19160]. Robustness to model mismatch is equally central: MRF-DiPh identifies motion and \(B_0/B_1\) inhomogeneity as unaddressed sources of mismatch, TPIDM shows that mis-specified physics can degrade performance on some datasets, and SA-PINN-ACTO confirms the expected difficulty of PINNs near sharp gradients and discontinuities [2506.23311]. A plausible implication is that future progress will depend less on a single “physics-informed diffusion” recipe than on principled choices about where physics should enter—loss, sampler, latent manifold, architecture, or reward—and how tightly it should constrain a learned generative prior.

Source: https://www.emergentmind.com/topics/physics-informed-diffusion