Abstract
Partial differential equations (PDEs) underpin quantitative descriptionsacross the physical sciences and engineering, yet high-fidelity simulationremains a major computational bottleneck for many-query, real-time, and designtasks. Data-driven surrogates can be strikingly fast but are often unreliablewhen applied outside their training distribution. Here we introduce NeuralOperator Warm Starts (NOWS), a hybrid strategy that harnesses learned solutionoperators to accelerate classical iterative solvers by producing high-qualityinitial guesses for Krylov methods such as conjugate gradient and GMRES. NOWSleaves existing discretizations and solver infrastructures intact, integratingseamlessly with finite-difference, finite-element, isogeometric analysis,finite volume method, etc. Across our benchmarks, the learned initializationconsistently reduces iteration counts and end-to-end runtime, resulting in areduction of the computational time of up to 90 %, while preserving thestability and convergence guarantees of the underlying numerical algorithms. Bycombining the rapid inference of neural operators with the rigor of traditionalsolvers, NOWS provides a practical and trustworthy approach to acceleratehigh-fidelity PDE simulations.