I do not know what the current state is of developing the ML target for refining against twinned (I, F) data is.

Formulas are there, see "Maximum likelihood refinement for twinned structures" here:

someone just need to code it. I might introduce it to phenix.refine some day.

Pavel