Accelerating convergence of a natural convection solver by continuous data assimilation

Abstract: The Picard iteration for the Boussinesq model of natural convection can be an attractive solver because it stably decouples the fluid equations from the temperature equation (for contrast, the Newton iteration does not stably decouple). However, the convergence of Picard for this system is only linear and slows as the Rayleigh number increases, eventually failing for even moderately sized Rayleigh numbers. We consider this solver in the setting where sparse solution data is available, e.g. from data measurements or solution observables, and enhance Picard by incorporating the data into the iteration using a continuous data assimilation (CDA) approach. We prove that our approach scales the linear convergence rate by $H^{1/2}$, where $H$ is the characteristic spacing of the measurement locations. This implies that when Picard is converging, CDA will accelerate convergence, and when Picard is not converging, CDA (with enough data) will enable convergence. In the case of noisy data, we prove that the linear convergence rate of the nonlinear residual is similarly scaled by $H^{1/2}$ but the accuracy is limited by the accuracy of the data. Several numerical tests illustrate the effectiveness of the proposed method, including when the data is noisy. These tests show that CDA style nudging adapted to an iteration (instead of a time stepping scheme) enables convergence at much higher $Ra$.