Generative diffusion models have revolutionized the synthesis of complex data, from images to molecular structures. However, in scientific domains, these models face a fundamental challenge: they are "physics-blind." A model trained on fluid dynamics might generate visually convincing turbulence that nonetheless violates the conservation of mass, while a protein generator might produce structures with impossible atomic overlaps.
Physics-guided diffusion sampling addresses this by treating physical laws not just as training data, but as active constraints during the generative process. By integrating differential equations, energy functions, and sensor measurements into the reverse diffusion steps, researchers can "steer" the model toward solutions that are both statistically likely and physically valid.
The Score Function as a Physical Compass
At its core, a diffusion model learns a score function—the gradient of the log-probability density of the data. During sampling, the model follows this gradient to transform noise into a clean sample. Physics-guided sampling augments this process by adding a second gradient term derived from physical residuals.
As Saydemir et al. demonstrate in the context of 4D turbulent flows, the residuals of the Navier-Stokes equations can act as a scoring mechanism. By calculating spatial and temporal derivatives during the denoising steps, the model nudges the evolving simulation toward states that satisfy fluid dynamics. Similarly, Jiahe Huang et al. use physical residuals to solve partial differential equations (PDEs) under partial observation, ensuring that the generated fields remain consistent with underlying laws even in regions where data is missing.
This approach works because the reverse diffusion process is iterative. Each step provides an opportunity to correct "hallucinations"—non-physical artifacts—before they become baked into the final output.
Grounding Generation in Data Consistency
In inverse problems like MRI reconstruction or seismic imaging, the goal is to find a high-resolution signal that matches a set of physical measurements. Here, physics guidance takes the form of data consistency.
Cui et al. show that by formulating MRI reconstruction as a model-driven stochastic differential equation (SDE), the drift term can be explicitly guided by the "self-consistency" of k-space (frequency domain) data. This ensures the model doesn't just generate a "plausible-looking" brain scan, but one that is strictly anchored to the raw sensor data. Aviles-Rivero et al. take this further by replacing standard Gaussian noise with k-space undersampling, treating the physical degradation process itself as the diffusion mechanism.
By alternating between a generative denoising step and a data-fidelity update, these frameworks prevent the model from diverging into physically impossible territory. Bangun and Cao describe this as a hybrid approach: the generative prior provides the structural "intelligence," while the physics-based optimization provides the "ground truth" anchor.
Steering with Differentiable and Non-Differentiable Oracles
Physics guidance is not limited to simple equations; it can incorporate complex dynamical constraints and even external software.
- Dynamical Constraints: Komorowska and Mathis integrate Normal Mode Analysis (NMA) into the sampling process to design proteins with specific movements, such as a hinge opening. Because NMA is differentiable, its gradients can steer the diffusion process without retraining the base model.
- Geometric Awareness: In robotics, Zhong et al. and Tao Zhong et al. use differentiable physics layers to guide dexterous grasping. By calculating gradients for stability and collision avoidance, the model generates hand configurations that are physically capable of holding an object.
- Non-Differentiable Guidance: When the physical law is locked inside a "black box" (like a quantum chemistry oracle), Shen and Zhang employ zeroth-order optimization to estimate gradients. This allows the model to leverage established physical laws to correct geometric distortions in molecular generation even when direct gradients are unavailable.
Beyond Gaussian Noise: Informative Bridges
While standard diffusion models start from unstructured noise, recent work suggests that starting from a physically informed "bridge" can be more efficient. Wu et al. introduce Informative Prior Bridges, which use Lyapunov functions to inject physical forces (like bond lengths) directly into the transition paths.
Similarly, Petrović et al. propose Curly Flow Matching, which uses a non-zero drift reference process to capture periodic physical dynamics, such as ocean currents or cell cycles. By replacing "straight-line" noise-to-data paths with paths that respect local directional priors, these models capture the "curly" realities of the physical world that standard diffusion might miss.
Why It Works: The Manifold of Plausibility
The success of these methods stems from a shared insight: the space of all possible data (e.g., all possible 3D grids) is vast, but the space of physically valid data is a much smaller, lower-dimensional manifold.
Purely data-driven models struggle to stay on this manifold because they only learn from examples. Physics-guided sampling provides the "restoring force" that pushes the model back onto the manifold whenever it drifts. Whether through the energy-based priors in TimeDiT or the virtual observations in Gao et al., this guidance ensures that the generative power of diffusion is harnessed within the rigorous boundaries of the physical world.
Key Foundational Work
A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction (2022) — contributed 2 concept elements to this direction.
PhysDiff: Physics-Guided Human Motion Diffusion Model (2023) — contributed 1 concept elements to this direction.
Diffusion Models for Phase Retrieval in Computational Imaging (2023) — contributed 1 concept elements to this direction.