|Our old two lattices seem to have converged for now. Looking at the best scores of the linac lattices so far, there's a trend for the "c2" versions to do better (and the "d2" versions to do worse) with the ones that are neither being in between.|
This suggests the "c2" parametrisation of the RF phase, where the 000-999 range corresponds to two cycles rather than one, is helping (but not the "d2", which is two cycles starting 180 degrees off). So I'm trying it with 7 and 9 linac blocks to see if we can beat the good performance of the "6Xc2" lattice (the best with v4.45).
Here is a picture of that RF phase stuff, which should make it clear.
A side effect is that I'm retiring the legacy Linac900Ext10tc2 lattice, which is still runnable with v4.44d. You'll all have to upgrade to v4.45 at this point (see improvements here).
|It's also interesting to note that 4.44d results dominate the top. In doing the best designs for past lattices, I've noticed that the 4.45 scores are always lower than the 4.44d results. It's the noise from the client, eliminated in 4.45, that gives those results that extra 'edge'.|
|And here's my brief write-up of it.|
|I've done a slightly more detailed comparison of the evolution of these linac lattices. The two graphs below plot the evolution of muon yield against number of simulations for all the linac lattices that have yields defined in comparable ways. The X-axis of results is actually a square-root scale so the behaviour early on isn't too bunched up near zero on the left.|
Coloured by RF phase parametrisation (Plain/c2/d2 as in the above post):
Coloured by number of linac blocks (1,3,6,7,8,9,10,12):
You can see that early on in the evolution, fewer block lattices evolve faster (turquoise and blue near the top) and many-block lattices evolve slower (red/orange near the bottom). Later on, it becomes less clear which one eventually ends up higher. I've done a lot of 6-block lattice because six seems to work well but it's not conclusive.