Multiscale modelling is a really interesting scientific challenge that is important in a number of areas. Basically, the issue is how to create models of systems where activities and interactions at a large number of different (temporal and/or spatial) activities happen at the same time. Due to computational costs and complexity constraints we cannot just model everything at the smallest level; yet, sometimes, small details matter.
I wonder if there is a role for some kind of machine learning here? This is a very vague thought, but I wonder if somehow we can use learning to abstract simple models from more detailed models, and use those simple models as proxies for the more detailed model, with the option to drop back into the detailed model when and only when it is specifically needed?