Stochastic partial differential equations (SPDEs) model physical systems under uncertainty, but applying them in practice usually requires knowing the magnitude and structure of the model uncertainty, which is often unknown and hard to infer from noisy measurements. A new paper on arXiv introduces the Stochastic Operator Network (SON), a framework that learns directly from noisy data and outputs both a mean solution field and a quantification of uncertainty.

SON combines the architecture of Deep Operator Networks (DeepONet) with Stochastic Neural Networks (SNNs) to model stochasticity and enable probabilistic prediction. The training procedure minimizes a Hamiltonian-type loss, with the objective optimized using the Stochastic Maximum Principle. This approach lets the network infer uncertainty from data rather than relying on a user-specified uncertainty model.

Numerical experiments on benchmark SPDEs under multiple uncertainty sources demonstrate the method's accuracy and robustness in capturing solution structure and quantifying predictive uncertainty. The authors report that SON performs well across these tests, suggesting it could be a practical tool for uncertainty quantification in complex physical models where noise and model error are poorly characterized.