arXiv:2611.23748v1 Announce Type: new Abstract: Reactive transport in heterogeneous marine sediments couples flow and multicomponent transport with nonlinear chemistry and evolving porosity and permeability, and its simulation over thousands of years is computationally expensive, especially when uncertain heterogeneity requires many simulations. We develop Split-and-Learn, a framework that keeps the sequential operator-splitting scheme of the full-order model and replaces each stage by a data-driven surrogate with its own structure. Transport is learned on a local data-driven multiscale space, either as a compressed reduced solution map (global inverse atoms) or as coarse-block pieces of the forward operator assembled into a Galerkin or least-squares Petrov-Galerkin system (local atoms). Chemistry is learned through bounded reaction extents with stoichiometric reconstruction, which preserves reaction invariants and nonnegativity. Flow uses learned interface traces and a reduced mixed multiscale solve that enforces coarse-block mass balance. We prove the stability of the split scheme with the evolving medium and bound the error of the learned solver by the accumulated one-step errors of its stages. Numerical experiments in two- and three-dimensional heterogeneous sediment domains show the accuracy and computational cost of the individual and combined learned stages against the full order split solver.