What is the deal with automatic differentiation

[gadiLhv]

I'm noticing a few FDTD libraries now supporting automatic differentiation, e.g. MEEP and Luminescent. From what I've read it is a method that avoids numerically calculating the derivatives, by updating them iteratively and storing multiple data-sets in the memory.

From what I've read, this method is implemented to enable topology optimization. Naturally I don't really get the relation, because to the best of my understanding, the medium parameters control the behavior of the fields, and not the other way around.

So what are the advantages of this method, if any?