Paper ReviewMathematics & StatisticsSimulation & Agent-Based

Score-Based Diffusion for Stochastic PDEs: When Generative Models Learn Physics Under Uncertainty

Stochastic PDEs model physical systems under uncertaintyโ€”from turbulent flows to financial markets. Score-based diffusion models, originally designed for image generation, are being repurposed to learn SPDE solutions adaptively, enabling Bayesian inference over complex physical systems without traditional MCMC.

By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

Score-based diffusion models have conquered image generation. They produce photorealistic images by learning to reverse a noise-corruption processโ€”starting from pure noise and iteratively removing it to reveal structured content. The mathematical core of this process is a stochastic differential equation (SDE) that defines the noise schedule, and a learned "score function" that guides the reverse process toward the data distribution.

This connection between diffusion models and SDEs is not merely a mathematical convenienceโ€”it is a bridge between generative AI and scientific computing under uncertainty. Stochastic partial differential equations (SPDEs) are the mathematical language for modeling physical systems where uncertainty is intrinsic: turbulent fluid dynamics, porous media flow, weather and climate, population dynamics, and financial markets. Solving SPDEs is computationally expensive because it requires sampling from high-dimensional probability distributions over space-time fields.

Huynh et al. propose using score-based diffusion models to learn SPDE solutions adaptivelyโ€”building a generative model that can produce samples from the solution distribution at any time step, updated as new observations arrive. This converts the SPDE solution problem into a sequential generative modeling problem, leveraging the same machinery that generates images to generate physics.

The Framework: Recursive Bayesian Inference via Diffusion

The system operates within a recursive Bayesian inference framework:

  • Prior: The current belief about the SPDE solution (a probability distribution over spatial fields)
  • Observation: New measurement data arrives (sensor readings, satellite observations, market data)
  • Update: Bayesian inference updates the belief given the observation
  • Prediction: The updated model predicts the solution at the next time step
  • Traditional approaches (Ensemble Kalman Filter, particle filters) implement this loop with explicit sample sets that become computationally prohibitive for high-dimensional SPDEs. Huynh et al. replace the explicit sample set with a score-based diffusion model that implicitly represents the solution distributionโ€”and can generate samples on demand.

    The advantage is that the diffusion model's representational capacity scales gracefully with the dimension of the solution space. A traditional particle filter with N particles provides O(N) resolution of the distribution; a score-based model with a fixed-size neural network can represent arbitrarily complex distributions.

    Financial Time Series: Geometry Meets Generation

    Kim et al. apply a related insight to a specific domain: financial time series generation via geometric Brownian motion (GBM). Standard diffusion models treat price trajectories as generic numerical sequences and add isotropic Gaussian noise during the forward process. This ignores the domain knowledge that stock prices are positive, multiplicative, and well-described by GBMโ€”the foundation of the Black-Scholes option pricing theory.

    By using GBM as the forward noising process (rather than isotropic Gaussian noise), the model incorporates financial domain knowledge directly into its generative process. The result: generated price trajectories that better capture the stylized facts of real financial dataโ€”heavy tails, volatility clustering, mean reversionโ€”because the generative process is geometrically appropriate for the data domain.

    This is a specific instance of a general principle: matching the SDE of the diffusion process to the SDE of the physical system produces better generative models than using generic noise processes. The principle extends to any domain where the data-generating process is governed by known SDEs.

    Claims and Evidence

    <
    ClaimEvidenceVerdict
    Score-based diffusion can learn SPDE solutionsHuynh et al. demonstrate framework; experimental validation on standard SPDEsโœ… Supported
    Diffusion models scale better than particle filters for high-dimensional SPDEsRepresentational capacity argument; comparative experiments limitedโš ๏ธ Theoretically motivated
    Domain-specific SDEs improve generative model qualityKim et al. show GBM-based diffusion captures financial stylized facts betterโœ… Supported
    The approach enables real-time Bayesian updatingAdaptive learning demonstrated; "real-time" depends on model size and hardwareโš ๏ธ Feasible for moderate dimensions
    Score-based methods replace traditional SPDE solversComplementary rather than replacement; traditional solvers provide different guaranteesโš ๏ธ Complementary, not replacement

    Open Questions

  • Error quantification: Traditional SPDE solvers provide error bounds. How do we quantify the error of score-based SPDE solutions? The generative model produces samples, but without guarantees on their fidelity to the true solution distribution.
  • Conservation laws: Physical systems obey conservation laws (mass, energy, momentum). Do score-based solutions respect these constraints, or must conservation be imposed as an additional loss term?
  • Multi-scale SPDEs: Many physical systems exhibit dynamics at multiple scales (turbulence, multiphase flow). Can score-based models capture multi-scale structure, or do they smooth over fine-scale dynamics?
  • Training data requirements: Training the diffusion model requires solution samples from the SPDEโ€”which must be generated by traditional solvers. How many training samples are needed, and does the computational cost of generating training data offset the efficiency gains of the learned model?
  • Transfer across parameter regimes: An SPDE model trained for one set of physical parameters (viscosity, diffusivity) may not generalize to different parameters. How do we build parameter-conditional score models that transfer across regimes?
  • What This Means for Your Research

    For computational mathematicians, the connection between generative AI and SPDE solutions is a bridge between two communities that have historically developed independently. The diffusion model community brings scalable architecture design and efficient sampling algorithms; the SPDE community brings rigorous problem formulations and error analysis frameworks.

    For financial engineers, the GBM-diffusion model (Kim et al.) provides a generative framework for financial time series that is both theoretically grounded (respecting the geometric structure of price processes) and practically flexible (generating diverse scenarios for risk management and stress testing).

    For the broader scientific computing community, score-based approaches to uncertainty quantification offer a path beyond the "curse of dimensionality" that limits traditional Monte Carlo methods for high-dimensional stochastic systems. The approach is nascentโ€”but the potential for computational efficiency gains in climate modeling, materials science, and fluid dynamics is substantial.

    References (2)

    [1] Huynh, T., Fajardo, R., Zhang, G. (2025). A Score-based Diffusion Model Approach for Adaptive Learning of Stochastic Partial Differential Equation Solutions. arXiv:2508.06834.
    [2] Kim, G., Choi, S., Kim, Y. (2025). A diffusion-based generative model for financial time series via geometric Brownian motion. arXiv:2507.19003.

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 8 keywords โ†’