Oh I didn't realise that there were any other approaches possible other than simulated annealing? For N=11 I can almost reach the optimal, but for N=12 my score is the same.
Yes, my previous approach for N>10 was a semiconstructional approach with an "ideal" borderless grid (dobrichev gives an example) as model. In this way you can get more than 24.7 points without extensive computation.
But yes, I think simulated annealing is a good path to follow and I will see how it works for larger grids. Meanwhile I reached 1084 for N=12 in this way. A central role is the definition of the "energy" of the grid and how "energy" changes if two cells are swapped. I suppose there are different possibilities here and the method used may have an impact on the quality of the solutions.