Hi Gerwald,

I just started using phenix.refine and it's working quite well so far. I 
ran the optimize weight strategy at the end of a test refinement and it 
lowered the Rfree as intended. Unfortunately, when I ran the next refine 
job based on the output .def file, it seems that the 'best weights' 
found in the previous run were not conserved as the geometry and R 
factors got a lot worse.

The "best weights" are good per macro-cycle. That is what is the "best weight" for macro-cycle #1 is not necessarily the best one for macro-cycle #2, etc. This is why they are not preserved for next runs.

Where can I find the weights determined by the optimize run and how can 
I fix those for further refinement?

If you run phenix.refine with "optimize_wxc=true", then you will see something like this:

================================ xyz refinement ===============================

Start r_work= 0.2359 r_free= 0.2576 bonds= 0.004 angles= 1.08
      r_work= 0.2304 r_free= 0.2504 bonds= 0.003 angles= 1.14 scale= 1.00
      r_work= 0.2286 r_free= 0.2476 bonds= 0.003 angles= 1.22 scale= 1.40
      r_work= 0.2279 r_free= 0.2465 bonds= 0.004 angles= 1.26 scale= 1.80
      r_work= 0.2271 r_free= 0.2454 bonds= 0.004 angles= 1.32 scale= 2.20
      r_work= 0.2266 r_free= 0.2447 bonds= 0.005 angles= 1.35 scale= 2.60
      r_work= 0.2262 r_free= 0.2442 bonds= 0.005 angles= 1.39 scale= 3.00
      r_work= 0.2258 r_free= 0.2436 bonds= 0.005 angles= 1.41 scale= 3.40
      r_work= 0.2256 r_free= 0.2433 bonds= 0.006 angles= 1.44 scale= 3.80
      r_work= 0.2253 r_free= 0.2429 bonds= 0.006 angles= 1.47 scale= 4.20
      r_work= 0.2251 r_free= 0.2426 bonds= 0.006 angles= 1.49 scale= 4.60
      r_work= 0.2249 r_free= 0.2425 bonds= 0.007 angles= 1.51 scale= 5.00
      r_work= 0.2245 r_free= 0.2417 bonds= 0.008 angles= 1.56 scale= 6.00
      r_work= 0.2242 r_free= 0.2420 bonds= 0.008 angles= 1.60 scale= 7.00
      r_work= 0.2238 r_free= 0.2408 bonds= 0.009 angles= 1.63 scale= 8.00
      r_work= 0.2236 r_free= 0.2413 bonds= 0.010 angles= 1.68 scale= 9.00
      r_work= 0.2234 r_free= 0.2409 bonds= 0.010 angles= 1.71 scale= 10.00
      r_work= 0.2314 r_free= 0.2516 bonds= 0.003 angles= 1.11 scale= 0.85
      r_work= 0.2331 r_free= 0.2538 bonds= 0.003 angles= 1.05 scale= 0.70
      r_work= 0.2362 r_free= 0.2566 bonds= 0.003 angles= 0.98 scale= 0.55
Final r_work= 0.2238 r_free= 0.2408 bonds= 0.009 angles= 1.63

In this particular case, the best result is:

r_work= 0.2238 r_free= 0.2408 bonds= 0.009 angles= 1.63 scale= 8.00

meaning that if you want to fix it then "wxc_scale=8" keyword in your subsequent phenix.refine run (I should replace "scale" with "wxc_scale" to make it more obvious).

Pavel.