Solver Convergence Improvements #536
Conversation
gusto/solvers/linear_solvers.py
Outdated
| @@ -33,13 +33,19 @@ | |||
| class TimesteppingSolver(object, metaclass=ABCMeta): | |||
| """Base class for timestepping linear solvers for Gusto.""" | |||
|
|
|||
| def __init__(self, equations, alpha=0.5, solver_parameters=None, | |||
| overwrite_solver_parameters=False): | |||
| def __init__(self, equations, alpha=0.5, tau_u=None, tau_r=None, tau_t=None, | |||
There was a problem hiding this comment.
Could this be a dictionary of tau values, where the keys are variable names? Otherwise this is specific to the compressible Euler equations whereas it would ideally be agnostic to equation type...
The other alternative is to not include this as an argument here.
(The complication being that this only works for shallow-water currently but that's a separate thing!)
There was a problem hiding this comment.
I've switched to a dictionary, with defaulting to alpha if the dictionary has no entry for a given field
gusto/solvers/linear_solvers.py
Outdated
| quadrature_degree=None, solver_parameters=None, | ||
| overwrite_solver_parameters=False): | ||
| """ | ||
| Args: | ||
| equations (:class:`PrognosticEquation`): the model's equation. | ||
| alpha (float, optional): the semi-implicit off-centring factor. | ||
| Defaults to 0.5. A value of 1 is fully-implicit. | ||
| tau_u (float, optional): the semi-implicit relaxation parameter for |
There was a problem hiding this comment.
I don't have a problem with each of the tau values being a separate argument here, but you could also make this a dictionary if you decide to go down that route.
gusto/solvers/linear_solvers.py
Outdated
| @@ -52,6 +58,23 @@ def __init__(self, equations, alpha=0.5, solver_parameters=None, | |||
| self.dt = equations.domain.dt | |||
| self.alpha = alpha | |||
|
|
|||
| # Set relaxation parameters. If an alternative has not been given, set | |||
| # to semi-implicit off-centering factor | |||
| if tau_u is not None: | |||
There was a problem hiding this comment.
these could actually be fit on one line nicely, e.g.
self.tau_u = tau_u if tau_u is not None else alpha
| """ | ||
| for field_name in transported_field_names: | ||
| if field_name != 'u': | ||
| logger.info(f'SIQN: Zeroing Implicit forcing for {field_name}') |
There was a problem hiding this comment.
I think the SIQN in the logged message now needs to be Semi-Implicit Quasi Newton
tommbendall
added
enhancement
…ject/gusto into solver_improvements
Co-authored-by: Dr Jemma Shipton <j.shipton@exeter.ac.uk>
Co-authored-by: Dr Jemma Shipton <j.shipton@exeter.ac.uk>
Importing solver improvements we have in GungHo:
tau values for linear solver which accelerate convergence if set to 1. This has been shown to make the moist baroclinic in channel stable for 2x2 SIQN with tau_t and tau_r = 1. Description of why these work has been attached in
Tau_Values.pdf
Accelerate solver convergence by zeroing implicit forcing's for transport terms
These changes have been added on switches, so do not change KGOs.