Iterable access to solver results and possibility of feedback to solvers

Currently QuTiP solvers are pure batch jobs. One calls a function, e.g. mesolve, which then runs until the state has been propagated to all requested time steps and then the function returns the result.

There are a variety of use cases where having access to results as they are calculated can be useful:

• Implementing a progress bar
• Limiting how long a solver runs for
• Stopping a solver early if a steady state or other desired or undesired state is detected.
• Implementing adaptive solvers where parameters are changed depending on the current state

We’re currently refactoring the QuTiP solvers for QuTiP v5, so perhaps now is a good time to consider whether this feature will see enough use to be worth the effort.

Implementation options:

• The solver could call a callback function each time a new result becomes available.
• The solver could yield each time a new result becomes available.
• The solver could do something asynchronous (I think this is not worth supporting natively in QuTiP – either of the others would allow something asynchronous to be implemented by a QuTiP users themselves if they wanted).

Probably these are not mutually exclusive. To make porting easier, it should still be possible to call some solve function without a callback or yielding as one can do now.

Related issues:

• qucontrol/krotov#93

[hodgestar]

The new QuTiP solvers now support solver.step(self, t, args=args). which allows manually stepping the state until time t with the optional supplied args. I don’t think this quite addresses the full feedback possibilities, but I am noting the progress here.

[Ericgig]

Most of scipy’s solver are not “thread-safe”, but we will be able to use other solvers in the new version. We will be sure to add at least one which is thread safe.

The args=new_args refer to the args in [H0, [H1, f(t, args)]], in your case, you could use something like:

def f(t, args):
    return args['u'] 

We don’t have step interpolation with changing coefficient array. It could be added, but we will have to think about it. Using a mutable, as in your example, would create error with most ODE solver with variable step, so I would not recommend it.