Update ExpRK documentation #144
Conversation
docs/src/solvers/ode_solve.md
Outdated
| - `ETDRK4` - 4th order exponential-RK scheme. Fixed timestepping only. | ||
| - `HochOst4` - 4th order exponential-RK scheme with stiff order 4. Fixed | ||
| timestepping only. | ||
| - `Exprb32` - 3rd order adaptive Exponential-Rosenbrock scheme. | ||
| - `Exprb43` - 4th order adaptive Exponential-Rosenbrock scheme. | ||
| - `Exprb32` - 3rd order adaptive Exponential-Rosenbrock scheme (broken at the moment). |
There was a problem hiding this comment.
Label it `"In development"
docs/src/solvers/ode_solve.md
Outdated
| @@ -95,6 +95,22 @@ parabolic PDEs, and thus it's suggested you replace it with `Rodas4P` in those | |||
| situations which is 3rd order. `ROS3P` is only third order and achieves 3rd order | |||
| on such problems and can thus be more efficient in this case. | |||
|
|
|||
| #### Exponential Integrators for Stiff Problems | |||
|
|
|||
| If explicit methods are preferred, then the exponential integrators provide an alternative | |||
There was a problem hiding this comment.
This section doesn't seem to be a high level description of the cases where it's useful and benchmarks well vs the other algorithms
docs/src/solvers/split_ode_solve.md
Outdated
| methods can allow for lazy calculation of `expm(dt*A)*v` and similar entities, and thus improve | ||
| performance. To tell a solver to use Krylov methods, pass `krylov=true` to its constructor. You | ||
| methods can allow for lazy calculation of `exp(dt*A)*v` and similar entities, and thus improve | ||
| performance. The algorithms that support this are: `LawsonEuler`, `NorsettEuler`, `ETDRK2`, |
There was a problem hiding this comment.
just keep one list and mark them appropriately
|
For the high level description section: I think you're correct in that we shouldn't give suggestion to users when no benchmarking is done. I'll remove this temporarily and add back after benchmarking is done (this will be my main focus for my follow-up work after GSoC btw). Also in addition to Exprb, I tagged ETD2 as "in development" because it's multistep and hasn't been updated for a while. The multistep ETD methods require some structural planning and could be a good project for future contributors/GSoC students. |
| orthogonalization procedure (IOP) [^1]. Note that if the linear operator/jacobian is hermitian, | ||
| then the Lanczos algorithm will always be used and the IOP setting is ignored. | ||
|
|
||
| You're required to provide the jacobian associated with the problem for the methods to work. |
There was a problem hiding this comment.
I thought it could use the autodiff or finite diff Jacobian like any of the other Jacobian-based methods?
There was a problem hiding this comment.
Ah not at the moment, but I can definitely mimic the implicit solvers and add support for autodiff/finite diff (adding uf to the cache, etc).
There was a problem hiding this comment.
yes, I think it's a good idea to do it, and it shouldn't be difficult since a system for handling it is already in place.
Adds missing integrators and describes in detail the options for the ExpRK methods.