Marwin Segler@Phlogiston1 good points! What I still need to make up my mind about is to what extent the “best” routes need to be returned during search, or whether the search algorithm should just return a long list of routes, that can then be refined according to desired criteria. Maybe that gives more flexibility to explore the results. I would guess setting up the right multi-objective scoring upfront might be tricky, similar to algorithmic molecular design.
@Phlogiston1 2/3
Can PMI be reasonably predicted? Or is it too complex?
The point on obvious routes is an interesting one. For comp Chem where 10000s of molecules need to be screened this seems desirable, but for process much less so. In my experience, with certain kinds of models, the algorithms can also return non-obvious solutions, in particular for more exotic heterocycle formation.