Just for jokes, I did a bend in SonnetLite (I work for Sonnet). SonnetLite is free from
https://www.sonnetsoftware.com, no time limit, problem size limited to 16 MBytes, 4 ports and 3 dielectrics. Packed project file is attached. Either unzip it, or change the extension to .zon and do File->Import in SonnetLite.
The bend is 0.75mm wide on input and 1.0mm wide on output, on 0.635mm Alumina substrate. Current distribution at 4 GHz is attached. Note current goes to the inside of the bend as expected. Zero current at the sharp corner means there is very high E-field there, as suggested above.
SonnetLite has an option to synthesize a lumped SPICE model for structures that are small with respect to wavelength. Read Chpt 22 in the user's manual (included in the SonnetLite download) for details. Briefly, it takes data at two frequencies and synthesizes the RLC combination that matches the data at those frequencies. If the lumped model is valid, you will get the same lumped model if you synthesize at two more frequencies. Be careful here because lumped models are limited in validity. Don't go prancing off with a bad model! Technical details in J. C. Rautio, "Synthesis of Lumped Models from N-Port Scattering Parameter Data," IEEE Tran. Microwave Theory Tech., Vol. 42, No. 3, March 1994, pp. 535-537 available from IEEE Explore, or I will email a pdf on request. Result (in PSPICE format) is:
* Sonnet Data File
* From: em Version : 10.52
* From Project: SimpleBend
* Data File Written: 02/22/2005 15:22:33
* < HDATE 02/22/2005 15:16:11
* < MDATE 02/22/2005 15:22:24
* Spice Data
* Limits: C>0.01pF L<100.0nH R<1000.0Ohms K>0.01
* Analysis frequencies: 2000.0, 3000.0 MHz
.subckt SimpleBend_0 1 2 GND
C_C1 1 GND 0.034582pf
C_C2 1 2 0.57792pf
C_C3 2 GND 0.136983pf
L_L1 1 2 0.077828nh
.ends SimpleBend_0
* Analysis frequencies: 3000.0, 4000.0 MHz
.subckt SimpleBend_1 1 2 GND
C_C1 1 GND 0.036716pf
C_C2 1 2 0.565649pf
C_C3 2 GND 0.139006pf
L_L1 1 2 0.077852nh
.ends SimpleBend_1
Looking at the differences between the two SPICE models, we can guess that the models are accurate to about 3% or so up to 4 GHz (the highest frequency analyzed for this example).
Note that all is as expected except...the port 1 to 2 connection is a parallel LC, not just an L. Interesting! As to why this is, it is undoubtedly due to the interaction of evanescent fringing fields as discussed above in this thread, but a more complete justification specific to this particular geometry I will leave to discussion. (i.e., I don't know why!)