SPDE (Stochastic Partial Differential Equation)

A stochastic partial differential equation (SPDE) is a partial differential equation augmented with stochastic forcing terms, typically modeled as noise processes, that govern the evolution of random fields over both space and time. Where ordinary PDEs describe deterministic continuum systems, SPDEs extend this framework to capture inherent randomness, measurement uncertainty, or unresolved fine-scale dynamics distributed across a spatial domain. Mathematically, they generalize stochastic differential equations (SDEs) to infinite-dimensional state spaces, requiring formulations in Hilbert or Banach spaces and careful interpretation choices — such as Itô versus Stratonovich conventions — along with specialized tools like semigroup theory, stochastic calculus in function spaces, and, for singular nonlinear cases, Martin Hairer's theory of regularity structures.

In machine learning, SPDEs have become practically important through several converging threads. One influential connection is the SPDE approach to Gaussian processes, where Matérn-class covariance functions can be represented as solutions to specific linear SPDEs, enabling scalable spatial statistics and structured priors over functions on irregular domains. More recently, the theoretical underpinnings of score-based and diffusion generative models are naturally expressed as SPDEs in the continuum limit, linking modern deep generative modeling to classical stochastic analysis. SPDEs also appear in physics-informed machine learning, where stochastic variants of governing equations allow neural networks to quantify uncertainty in simulations of fluid dynamics, climate systems, and other spatially extended phenomena.

The practical relevance of SPDEs to ML has grown substantially since the 2010s, driven by advances in probabilistic numerics, Bayesian inverse problems constrained by PDEs, and the rise of diffusion-based generative models. Numerically solving SPDEs remains challenging, motivating multilevel Monte Carlo methods, structure-exploiting discretizations, and learned surrogates that approximate SPDE solutions at a fraction of classical computational cost. As ML increasingly targets scientific and physical domains where spatiotemporal uncertainty is fundamental, SPDEs provide a rigorous mathematical language connecting probabilistic modeling, generative learning, and physics-based simulation.