Physics-informed neural networks typically assume deterministic governing equations. A new paper introduces stochastic physics-informed neural networks (SPINNs) to handle stochastic boundary-value problems where the operators themselves are random. The key idea is a structural probabilistic coupling: a shared Gaussian component links the network outputs, so the model respects the underlying physics while accounting for randomness.
The approach is developed for compact boundaryless multiply connected manifolds, a class of geometries that goes beyond the simple domains often considered in PINN research. This makes the method relevant for problems where the solution domain has holes or nontrivial topology.
The authors position SPINNs as a tool for random-operator models, with applications in machine learning and uncertainty quantification. By embedding stochasticity directly into the network architecture, the framework offers a principled way to quantify uncertainty in solutions to boundary-value problems.