usx, what was your model, what Er value did you used in ADS? Only sections (2)-(1)-(5)?
Here is my recent result from finite element code run with pretty dense grid
parameters: er=3.66 , h=0.25 mm (ideal substrate, PEC + PML)
frequency = 24.000 GHz
Z0=50;
S11=0.306-0.065*i;
Z11=Z0*(1+S11)/(1-S11)
Z11 = 92.8- 13.4*i
port 1 (before taper) is set to 50 ohms, and it sees 92.8 Ohm. not sure if it is correct. At least this code gives good values for single patch. But metal layer i use is very thick.
I think fastest solution is to make prototype board with stepped sizes (5 steps).
Or use other matching structure, for example open-ended stub, which could be tuned by cutting. And after good result is obtained redo the simulation.
From simulation i have noticed that tapered part (5) is kind of sensitive to ground metal near ic, and maybe answer is that it more like a grounded coplanar waveguide at ic end.