Rana suggested including some additional terms to the cost function to penalize high sensitivity to deviations in the layer thickness (L). So the list of terms contributing to the cost function now reads:
 Thermal noise  we use the proxy function from E0900068v3 to do this
 Deviation from target T @1064nm, ppol
 Deviation from target T @532nm, p and spol
 HR Surface field
 The ratio with dL/L = 1%, evaluated at 1064nm ppol and 532nm p and spol (only the latter two for the AR side)
I did not include other sensitivity terms, like sensitivity to the refractive index values for the low and high index materials (which are just taken from GWINC).
There is still some arbitrariness in how I chose to weight the relative contributions to the cost function, but after some playing around, I think I have a solution that I think will work. Here are the spectral reflectivity and layer thickness plots for the HR and AR sides respectively.
HR side: for a 1% increase in the thickness of all layers, the transmission changes by 5% @ 1064nm ppol and 0.5% @ 532nm s and ppol
AR side: for a 1% change in the thickness of all layers, the transmission changes by <0.5% @ 532nm s and ppol
(substrate to the right of layer 38)
I've also checked that we need 19 layer pairs to meet the spec requirements, running the code with fewer layer pairs leads to (in particular) large deviations from the target value of 50ppm @ 1064nm ppol.
Do these look reasonable?
