1: C# Library with interfaces and/or abstract base classes
2: F# library with implementations of those interfaces and base classes
3: C# program (console, web service, GUI, ect) that specifies the implementations in Dependency Injection
Or is there a simpler way for C# and F# to co-exist in the same project (dll or exe)?
For that matter, you don't even need the interfaces if you wouldn't have had them in a C#-only solution. Just define the class in F# and use it directly from C#.
You still need a separate assembly for F#, but that doesn't imply dependency injection - again, just reference it and use it.
The basic decision making logic needs to be very simple and easy to follow by scientists and people who aren't day-to-day software engineers: Basically, input some data, such as sensor readings, run it through an algorithm that makes decisions, and then output the decisions.
The biggest advantage of F# is its gradual typing and full type inference which allows to massively reduce the amount of text required to describe application or domain logic. It is also extremely composable and I find doing async in F# somewhat nicer than in C# too. F# also has better nullability (or, rather, lack of thereof) assurances and, in my opinion, better UX for records.