Solving partial differential equations often forces a trade-off: classical numerical solvers are accurate but computationally expensive, while machine-learning methods are fast but can suffer from spectral bias, failing to capture high-frequency components. One way around this is a hybrid iterative solver that, at each iteration, selects from an ensemble of solvers to exploit their complementary strengths. However, choosing the optimal sequence of solvers is a challenging combinatorial problem.
Greedy selection is attractive because it comes with a constant-factor approximation guarantee under Lipschitz assumptions, but it requires knowledge of the true error at every step—information that is not available in practice. The authors address this by proposing an approximate greedy router that efficiently mimics the greedy approach to solver selection without relying on true-error feedback.
In experiments on the Poisson, convection-diffusion, and anisotropic diffusion equations, the method consistently reduced time-to-solution relative to single-solver baselines and to existing hybrid approaches such as HINTS. The paper is a single-source contribution, so no conflicting findings are noted.