Fixed-Time Solvers
Fixed-Time Solvers
If solve! is applied to a PDEDescription and a FEVector (that specifies the ansatz spaces for the unknowns) an investigation of the PDEDescription is performed that decides if the problem is nonlinear (and has to be solved by a fixed-point algorithm) or if it can be solved directly in one step. Additionally the user can manually trigger subiterations that splits the fixed-point algorithm into substeps where only subsets of the PDE equations are solved together.
CommonSolve.solve!
— Functionfunction solve!(
Target::FEVector, # contains initial guess and final solution after solve
PDE::PDEDescription;
kwargs)
Solves a given PDE (provided as a PDEDescription) and writes the solution into the FEVector Target (which knows the discrete ansatz spaces).
Keyword arguments:
anderson_iterations: use Anderson acceleration with this many previous iterates (to hopefully speed up/enable convergence of fixpoint iterations). Default: 0
subiterations: an array of equation subsets (each an array) that should be solved together in each fixpoint iteration. Default: ''auto''
showiterationdetails: show details (residuals etc.) of each iteration. Default: true
anderson_unknowns: an array of unknown numbers that should be included in the Anderson acceleration. Default: [1]
reltol: reltol for linear solver (if iterative). Default: 1.0e-12
show_statistics: show some statistics like assembly times. Default: false
anderson_metric: String that encodes the desired convergence metric for the Anderson acceleration (possible values: ''l2'' or ''L2'' or ''H1''). Default: ''l2''
skipupdate: matrix update (for the j-th sub-iteration) will be performed each skipupdate[j] iteration; -1 means only in the first iteration. Default: [1]
linsolver: any AbstractFactorization from LinearSolve.jl (default = UMFPACKFactorization). Default: LinearSolve.UMFPACKFactorization(true, true)
damping: damp the new iteration with this part of the old iteration (0 = undamped), also a function is allowed with the interface (olditerate, newiterate, fixed_dofs) that returns the new damping value. Default: 0
time: time at which time-dependent data functions are evaluated or initial time for TimeControlSolver. Default: 0
parallel_storage: assemble storaged operators in parallel for loop. Default: false
showsolverconfig: show the complete solver configuration before starting to solve. Default: false
anderson_damping: Damping factor in Anderson acceleration (1 = undamped). Default: 1
abstol: abstol for linear solver (if iterative). Default: 1.0e-12
checknonlinearresidual: check the nonlinear residual in last nonlinear iteration (causes one more reassembly of nonlinear terms). Default: ''auto''
fixed_penalty: penalty that is used for the values of fixed degrees of freedom (e.g. by Dirichlet boundary data or global constraints). Default: 1.0e60
target_residual: stop fixpoint iterations if the absolute (nonlinear) residual is smaller than this number. Default: 1.0e-12
maxiterations: maximal number of nonlinear iterations (TimeControlSolver runs that many in each time step). Default: ''auto''
Depending on the subiterations and detected/configured nonlinearities the whole system is either solved directly in one step or via a fixed-point iteration.
Anderson acceleration
Fixpoint iterations my be accelerated by Anderson acceleration. Concepts and some theoretical background can be found in the reference below. Within this package, Anderson acceleration can be triggered by optional solver arguments: the user can specify the depth of the Anderson acceleration (anderson_iterations), the damping within the Anderson iteration (anderson_damping), the ids of the unknowns that should be included in the iteration (anderson_unknowns) and the convergence metric (anderson_metric); also see above for a full list of optional solver arguments. In case of subiterations, the Anderson iteration will be called as a postprocessing after the final subiteration.
Reference:
"A Proof That Anderson Acceleration Improves the Convergence Rate in Linearly Converging Fixed-Point Methods (But Not in Those Converging Quadratically)",
C. Evans, S. Pollock, L. Rebholz, and M. Xiao,
SIAM J. Numer. Anal., 58(1) (2020),
>Journal-Link<