📝 SummarXiv

Abstract

We present SRaFTE (Super-Resolution and Future Time Extrapolation), a two-phase learning framework that couples coarse grid solvers with neural operators to super-resolve and forecast fine grid dynamics for time-dependent partial differential equations (PDEs). In our construction, Phase 1 learns a super-resolution map from coarse to fine solutions, while Phase 2 embeds this operator in a predictor-corrector loop with the coarse solver, forming an operator composition that serves as a surrogate fine grid propagator for future-time extrapolation. We benchmark SRaFTE on three canonical two-dimensional PDEs of increasing dynamical complexity: the heat equation, the wave equation, and the incompressible Navier-Stokes equations. Compared to well-known benchmarks, SRaFTE provides reliable super-resolution in Phase 1 and delivers consistently accurate long-term forecasts in Phase 2 across all three examples for new test data. Our results suggest that coupling a learned super-resolution operator with a coarse grid solver provides an effective and efficient means of modeling high-resolution spatiotemporal dynamics, particularly when the dynamics of the PDEs at the coarse and fine resolutions exhibit pronounced scale separation.