Some things to think of:
1 - it must be hard to generate one block, but cheap to verify. I doubt that the simulations in this context can be easily verified.
2 - difficulty must be adjustable to keep block rate mostly constant. Of course, you could adjust the simulated time for each unit, don't know how well that would work.
3 - each block must depend on the previous block, i.e. it must not be possible to perform all or part of the hard work before the previous block is known.

I have a gut feeling that most grid computing efforts don't fit these constraints well.

Onkel Paul