Most model reduction methods are space-only in that they reduce the spatial dimension of the solution but not the temporal one. These methods integrate an encoding of the state of the nonlinear dynamical system forward in time. We propose a space-time method -- one that solves a system of algebraic equations for the encoding of the trajectory, i.e., the solution on a time interval $[0,T]$. The benefit of this approach is that with the same total number of degrees of freedom, a space-time encoding can leverage spatiotemporal correlations to represent the trajectory far more accurately than a space-only one. We use spectral proper orthogonal decomposition (SPOD) modes, a spatial basis at each temporal frequency tailored to the structures that appear at that frequency, to represent the trajectory. These modes have a number of properties that make them an ideal choice for space-time model reduction. We derive an algebraic system involving the SPOD coefficients that represent the solution, as well as the initial condition and the forcing. The online phase of the method consists of solving this system for the SPOD coefficients given the initial condition and forcing. We test the model on a Ginzburg-Landau system, a $1 + 1$ dimensional nonlinear PDE. We find that the proposed method is $\sim 2$ orders of magnitude more accurate than POD-Galerkin at the same number of modes and CPU time for all of our tests. In fact, the method is substantially more accurate even than the projection of the solution onto the POD modes, which is a lower bound for the error of any space-only Petrov-Galerkin method.
翻译:暂无翻译