Open hardware microwave vector network analyzer

[loxodes]

I'm developing an open hardware microwave network analyzer as a hobby project, I thought I'd post it here to increase my post count to qualify for the eevblog tequipment discount code share it with y'all.
This is first attempt at building test equipment, I welcome any feedback. I'm not expecting this to compete with $20,000 Keysight VNAs, I'll be happy with 60 dB or so of dynamic range up to 10 GHz with measurements within about .5 dB of reality. I'm still working on bringing up a prototype, currently only 1 port measurements from about 2 GHz to 5 GHz are mostly working. Over the next month or two I hope to have two port measurements between 2 GHz and 10 GHz.

This isn't the Henrik Forsten's network analyzer that you may have seen recently, he scooped me by a few months (http://hforsten.com/cheap-homemade-30-mhz-6-ghz-vector-network-analyzer.html). If you are looking for something to build, his is the better choice at the moment.

The VNA is a bog standard 4 receiver design, all of the hard work is done by MMICs and off-the-shelf modules. Here is a picture of my prototype, it a bit of a mess at the moment:


The VNA is controlled using a BeagleBone Green, with a PRU reading samples from four AD9864 IF digitiziers. An EMC nightmare of ribbon cables erupting off a custom shieldcape ties together everything. I'm considering spinning a custom board around an Octavo OSD335x once everything else is working.


The VNA works by injecting a signal into the device under test and comparing the injected signal with the reflected signal. This design has a non-zero IF, so it needs an RF and LO.
These are generated by frequency synthesizer boards. I built them around TI LMX2592 20 MHz to 9.8 GHz synthesizers with some extra circuity for variable attenuation and filter banks. The BOM cost of these is about $300 each, I'm sitting on a unbuilt redesign which should cut that to about $150. These could be useful as a low cost standalone microwave signal generators. Earlier revisions had a power detector and amplification that I took out to reduce complexity.. I may add those back in at some point..

Mistakes on the filter footprints limit the usefulness of these above about 6 GHz. I'll fix it in the next revision.. until then, I'm using an external multiplier. For the LO, the LTC5548 mixer has a built in multiplier. For the RF path, I'm planning on switching in an AMMP-6120.

The receivers are built around the LTC5548 mixer and AD9864 ADC. I designed them to also be useful for a Greg-Charvat-style through the wall radar, there is a crystal filter in the IF path.


So far only one port measurements are tested and they are still a little wobbly. Here is an example plot of S11 of a 10 dB attenuator after short/load/open one port calibration (ideally, this would be a flat S11 of -20 dB):



As y'all can see, it isn't working perfectly yet. I still have some work to do on the software, and some of the harmonic filtering for the synthesizers is bypassed due to fried pins until I order a fresh BeagleBone.


After I add in a multiplier and finish the software for two port measurements/calibration, this is expected to work from about 2 GHz to 10ish GHz. I plan to use this as a testbed for attempting more ambitious measurements such as triggered and pulsed s-parameters, cold source noise figure measurements, time domain analysis...


All the software and hardware is available on my github under an MIT license: https://github.com/loxodes/vna. I'll post occasional updates here as I get more things working.

[hendorog]

Awesome. 

One thing, the S11 attenuator measurement may not be perfect but might not be as bad as you think. I think the VSWR of the 10dB attenuator will affect the measurement.

I measured the S11 of a couple of 10dB attenuators on my 8753 (<3GHz) and they are about +/- 0.5dB. OTOH S21 is much flatter than S11.

I expect this is due to the VSWR spec being only slightly better than 1.2 - which is ~20dB return loss on its own.

[G0HZU]

Yes, a lot will depend on the quality of the SMA attenuator. However, the S11 plot looks realistic on the right hand side of the plot but the sharp transitions on the LHS of the plot do look a bit suspicious. So maybe something in the system is clashing and beating in and out of phase here? I wonder if you could be flexing the various cables as you do the SOL calibration here? You will need to use very (phase) stable cables to get the best out of the system. How much does the plot change if you wiggle things in the system? Maybe it all needs to be bolted/strapped down?

See below for the RL of a decent SMA 10dB attenuator up to 5GHz when the far end is left open. I measured this on an Agilent E5071 VNA. This is a Suhner 18GHz 10dB attenuator. I think If I tested a cheaper MiniCircuits (VAT10) SMA 10dB attenuator I'd expect to see lots of ripple in the S11 plot across LF thru 5GHz. Maybe a few dB  peak to peak? However, I'd expect the ripple to look smooth and regular without the spiky transitions seen in the LHS of that first plot

[loxodes]

Thanks for the help.

I was measuring an Omni-Spectra/Macom 2600, 18 GHz 10 dB attenuator.
The attenuator is old enough that I can't track down any specs on max VSWR. I'll sneak it into work some time and take some measurements on a real VNA to know what to expect.

My other measurements suggest that the attenuator isn't entirely to blame, here is S11 on a Mini-Circuits VAT-3+ that shows the same jump at 2.5 GHz:


The cables and front end modules are flopping around in a pile, I'll try securing things. Sweep sweep is currently only few points per second so looking at the effect of wiggle will be tricky. I should be able to plot a single frequency with a decent refresh rate, I'll try writing something for that.

Internal reflections within the front end are also likely, I designed this open loop without using a VNA so the switch, amplifiers, resistive splitters, and mixers may not be matched well. I'll investigate if/where adding internal attenuators helps.

I'm being sent into the wilderness for field work tomorrow morning, when I'm back next week I'll write up a block diagram to articulate the design better.

[Bud]

I did not study your schematic but if you have RF amplifiers in the path, you should check them with a wideband spectrum analyzer for oscillation.
Same for the front end mixer ICs if you have any. Connect the SA to the mixer RF input and see if any oscillation is observed at frequencies other than LO or its harmonics.
RF amplifiers may oscillate because of design or layout issues and well into GHz range, so you need a decent SA  to check for that. Mixer ICs can oscillate because they may be only conditionaly stable, i.e. stable within a certain range of impedances at their LO and/or RF inputs.
Apart from that, you probably should shield and mount the modules in some sort of enclosure so they do not flap in the breeze before doing anything else.

[G0HZU]

My other measurements suggest that the attenuator isn't entirely to blame, here is S11 on a Mini-Circuits VAT-3+ that shows the same jump at 2.5 GHz:

I have a VAT3 SMA attenuator here and did a quick look at a 1 port measurement of the attenuator at both the male and female ends. The male end plot has a basic port extension correction for a barrel adaptor included and you can see that the attenuator looks pretty much the same at either end if the ref plane is set at the SMA of the VAT3. This was a hasty measurement and normally I'd use a better correction than a basic port extension but the results look close enough anyway. The female end of the attenuator is measured directly with no port extension required.

You can see how much the real and imaginary values change across frequency. I was surprised to see it change so quickly as there would need to be a fair bit of tline length within the structure to get this. Almost the full length of the part? approx 100ps? I think my measurement is OK but I'll check it again tomorrow.

Hope this is useful.

[G0HZU]

I just checked my VNA results against the official MiniCircuits s parameter data for this VAT3+ attenuator. See below. It looks like my measurement is valid as the agreement is very close. I'm comparing my data against MiniCircuits in the plot below. The traces below almost overlap!

I'm using a E5071 4 port VNA and the 13.5GHz N4431-60006B Ecal module here.

This VAT3+ attenuator could prove to be a useful benchmark for you to measure against to see how well your system develops/improves over time. The MiniCircuits data looks to be good to me!

[loxodes]

This VAT3+ attenuator could prove to be a useful benchmark for you to measure against to see how well your system develops/improves over time. The MiniCircuits data looks to be good to me!

Good to know! I hadn't thought to check if MiniCircuits provided s-parameters and was wondering what I could use as a reference.

I did not study your schematic but if you have RF amplifiers in the path, you should check them with a wideband spectrum analyzer for oscillation.
Same for the front end mixer ICs if you have any. Connect the SA to the mixer RF input and see if any oscillation is observed at frequencies other than LO or its harmonics.

I poked at the amplifiers and mixers with a e-field probe and a SA-124B spectrum analyzer, and I can see why you recommended building an enclosure 

The following are sweeps from 10 MHz to 8 GHz, with the VNA measuring at 5 GHz.
I'm using the built-in doubler on the mixer, so this means a LO of 2.5225 GHz and a RF of 5 GHz.
The IF frequency is 45 MHz, the ADCs have built in mixers with a 48.25 MHz LO for a baseband frequency of 3.25 MHz. The synthesizer is running off a 100 MHz clock, the ADC clock is 26 MHz.

My initial guesses at the markers are:
1 - FM radio station
2 - FM radio station, coupled into amplifier, then doubled?
3 - LO/2 synthesized by LMX2592
4 - RF
5 - LO (doubled by mixer)
6 - This occurs even when the VNA is off. I think this is 5 GHz WiFi channel 44 or 46. My router is 2.4 GHz only so this may be a neighbor.
There are a bunch of garbage spurs in the spectrum up to 2 GHz that I need to investigate further.
Anyways, here are the plots:

Output of LO synth amplifier


LO input to mixer


RF input to mixer


RF synth amp output


One foot in the air above the table with the VNA turned on


One foot in the air above the table with the VNA turned off 


I should gather these at a few VNA frequencies and draw up a block diagram/high level schematic, but I need to pack for some field work..
It looks like synthesizing the LO directly and disabling the doubler built into the mixer may make things a little simpler.

Thanks, y'all have given me a lot to think about.

[Bud]

What is that last spur at 8.5GHz that gets 20dB bigger at the mixer RF input, this is worrying and may indicate oscillation. If you can cut supply power to the mixer try that, if the spur goes away you may have a problem with the mixer.

[dcarr]

First of all, awesome.  Very cool to see people experimenting at this level.

A small note on connectors:

I built some X band modules using very similar looking connectors to coplanar launch ones above.  I was surprised to find that the return loss of the SMA->CPWG transition wasn't all that great (~10-15dB IIRC).  Ultimately, I went with Johnson 142-0711-826 (with center pin cut down to about ~.06" long) and got better results.  I was using 0.062" FR4 PCBs from OSH Park and 32 mil center trace / 6 mil gaps.

Don't know if you'll encounter the same trouble I did, but it might we worth making a simple test board and taking some measurements.  Who do you use for your PCB supplier?  I also had trouble with OSH Park delivering boards with fairly variable dielectric constant which also made things challenging.

David

[loxodes]

What is that last spur at 8.5GHz that gets 20dB bigger at the mixer RF input, this is worrying and may indicate oscillation. If you can cut supply power to the mixer try that, if the spur goes away you may have a problem with the mixer.

The spur at 7.567 GHz was the third harmonic of LO/2. I ran out of markers on the spectrum analyzer and forgot to talk about it.. It looks like I need to pull out the mixer module and take some more measurements to verify that everything is working as I'm expecting. I don't see 3LO to RF leakage specified on the LTC5548 datasheet, but 2LO to RF isolation appears to be on the order of 20 dB to 25 dB. (My measurements were with a near field probe, so the amplitudes were relative. Assuming 20 dB 3LO to RF isolation, I'd expect 3LO to be about the same amplitude as the RF signal input into the mixer.)

The spur at 5.23 GHz may have been an artifact from the SA124B spectrum analyzer. It reliably disappears if I modify the sweep settings and it doesn't reduce in amplitude when I terminate the input of the spectrum analyzer.

Disabling the internal multiplier in the mixer removed the signal at ~7.567 GHz and the discontinuity near 2.5 GHz in S11 measurements. I'm still a dB or two off from MiniCircuits s-parameters for the VAT-3+, but I have plenty of potential problems to investigate.

A small note on connectors:

I built some X band modules using very similar looking connectors to coplanar launch ones above.  I was surprised to find that the return loss of the SMA->CPWG transition wasn't all that great (~10-15dB IIRC).  Ultimately, I went with Johnson 142-0711-826 (with center pin cut down to about ~.06" long) and got better results.  I was using 0.062" FR4 PCBs from OSH Park and 32 mil center trace / 6 mil gaps.

Don't know if you'll encounter the same trouble I did, but it might we worth making a simple test board and taking some measurements.  Who do you use for your PCB supplier?  I also had trouble with OSH Park delivering boards with fairly variable dielectric constant which also made things challenging.

David

Yeah, I've encountered the same trouble.. The Johnson 142-0771-831 is the forth SMA connector in twice as many footprints I've tried with OSH Park 4 layer for this project and I'm still not entirely happy with the results. I wish they offered RT4350B...

My current stackup for grounded CPW on OSH Park is a 1 mm wide trace on layer 1 with a 0.2 mm gap to the top layer ground plane, and a ground plane on layer 3. This gives an insertion loss through a 50 mm long test board of about 1-2 dB out to 10 GHz on a scalar network analyzer. Plugging the stackup into a coplanar waveguide calculator, about half of that loss might be from the trace and half from the connectors. I'm afraid to see what the return loss looks like, I should take it to work and see 

Thanks for the suggestion, I'll put a few 142-0711-826 connectors on my next digikey order.

[tipofthesowrd]

FYI: the return loss from a coaxial to planar transition is something I have also struggled with.
Not entirely applicable but there is an interesting read online as a sample from the Essentials of RF and Microwave Grounding book.
http://www.artechhouse.com/uploads/public/documents/chapters/holzman_941_ch04.pdf

They present the case where the coaxial (SMA) to microstrip transition is first modeled as usual with the top pads connected with vias to the bottem (microstrip ground plane) and you already see a deterioration in return loss above 5 GHz.

Afterwards they model it is as CPW to microstrip transition where they remove the ground plane under the SMA connector.

[henrikf]

Nice to see that other people are also working on making their own microwave test equipment.
I would have also wanted to make my VNA for higher frequencies, but the costs would have increased too much for my budget, but it seems that it isn't limiting you.

I don't quite understand how your VNA architecture works. Is the source connected only to one of the ports? It seems that the test ports are not identical. Is the power splitter working as a directional coupler?


I had a similar looking but less serious problem with my VNA's receiver that you are seeing that were caused be poor repetability and linearity of the receiver.

I used MAX2871 chip as the RF signal source for test signal and mixer LO. It has several internal VCOs that work at different frequencies and automatic logic for choosing the right one. It turns out that the different VCOs have different output power and output impedances that slightly change the receiver mixer's conversion gain. Only a fraction of a dB, but enough to be seen in the measured S-parameters. If different VCOs are chosen for the same frequency point when measuring different calibration standards, different gain of the receiver is seen as an error in the calibration coefficients. In practice it can been seen as a spike in the S-parameter plot at one frequency in all measurements.

Auto gain functionality in your ADC chip is one source of repeatibility errors and you should control the gain manually if you aren't already doing that.

Linearity of the receiver on my VNA is also not that good and the best measurements are obtained with slightly lower output power. It is partly caused by driving a balanced mixer single-endedly and partly because of a bad IF amplifier choice. You have much higher quality components and I don't see any obvious issues with linearity, but it might be worthwhile to try the measurement with a lower RF power to see if the accuracy improves.

How do the calibration error coefficients look like? They tell a lot about the RF performance of the VNA. If you are using scikit-rf for calibration you can access them with "cal.coefs".

[loxodes]

I would have also wanted to make my VNA for higher frequencies, but the costs would have increased too much for my budget, but it seems that it isn't limiting you.

Yes, your design is far more elegant.. I've spent at least $300 on just SMA connectors 

I don't quite understand how your VNA architecture works. Is the source connected only to one of the ports? It seems that the test ports are not identical. Is the power splitter working as a directional coupler?

Yeah, the power splitters approximate directional couplers. I intend to replace them with actual directional couplers eventually.. The annotated picture on my GitHub page is an earlier prototype inspired by "Design and implementation of a compact Vector Network Analyzer" by  Martínez Argudo. (I think this is a link to the paper, my network connection is not stable enough to download the paper and confirm that: https://riunet.upv.es/bitstream/handle/10251/34999/MartinezArgudo,%20Marcos_Movilidad_Abierto.pdf?sequence=1)

After reading about your project I decided to construct a similar VNA to avoid the requirement for a broadband quadrature demodulator, here is an updated diagram of the prototype pictured earlier in this thread. It is essentially a copy of the four receiver architecture presented on your blog:

I had a similar looking but less serious problem with my VNA's receiver that you are seeing that were caused be poor repetability and linearity of the receiver.

I used MAX2871 chip as the RF signal source for test signal and mixer LO. It has several internal VCOs that work at different frequencies and automatic logic for choosing the right one. It turns out that the different VCOs have different output power and output impedances that slightly change the receiver mixer's conversion gain. Only a fraction of a dB, but enough to be seen in the measured S-parameters. If different VCOs are chosen for the same frequency point when measuring different calibration standards, different gain of the receiver is seen as an error in the calibration coefficients. In practice it can been seen as a spike in the S-parameter plot at one frequency in all measurements.

Auto gain functionality in your ADC chip is one source of repeatibility errors and you should control the gain manually if you aren't already doing that.

Linearity of the receiver on my VNA is also not that good and the best measurements are obtained with slightly lower output power. It is partly caused by driving a balanced mixer single-endedly and partly because of a bad IF amplifier choice. You have much higher quality components and I don't see any obvious issues with linearity, but it might be worthwhile to try the measurement with a lower RF power to see if the accuracy improves.

How do the calibration error coefficients look like? They tell a lot about the RF performance of the VNA. If you are using scikit-rf for calibration you can access them with "cal.coefs".

Thanks for the insights! I'll investigate these next week once I'm back home.

[Bud]

The Peregrine switch isolation is insuficcient for the transmission mode, Consider using two of them in series in each arm, a total of 3 switch ICs. Even for the reflection mode it is poor above 3GHz. Also you may want to use reflective switches, not absorptive ones to further improve isolation.

With power splitters in place of directional couplers you technically should be using 6dB attenuators, not 16. This is what two-way loss of a power splitter is. You robbing yourself off of 10 dB dynamic range.

[henrikf]

I've spent at least $300 on just SMA connectors 

The SMA connector layout I used on my VNA with a ground cut under the center pad has a return loss better than -20 dB at 10 GHz according to the EM simulations. I haven't measured it that high, but I have used it in few boards working under 6 GHz and the return loss has been excellent. The SMA connector I am using (Digikey part no. CONSMA003.062-ND) is much cheaper than your connectors. You might want to try it on your boards if you want to save some money on the connectors.

Yeah, the power splitters approximate directional couplers. I intend to replace them with actual directional couplers eventually.. The annotated picture on my GitHub page is an earlier prototype inspired by "Design and implementation of a compact Vector Network Analyzer" by  Martínez Argudo. (I think this is a link to the paper, my network connection is not stable enough to download the paper and confirm that: https://riunet.upv.es/bitstream/handle/10251/34999/MartinezArgudo,%20Marcos_Movilidad_Abierto.pdf?sequence=1)

That architecture makes more sense.

There's a problem with using a power splitter as a directional coupler as you have drawn it. Directivity of the A/B reflection receivers is determined by the isolation of the power splitter and should be good enough, but the reference receiver RX1/RX2 directivity is not so good. In fact it seems that reference receiver directivity is decided by the insertion loss of the power splitter, which is only about 3 dB. Reference receiver directivity shouldn't be as important as the reflection receivers, but 3 dB seems very low. There are some equations on how it affects the error terms in this document pages 7 and 9: http://emlab.uiuc.edu/ece451/appnotes/Rytting_NAModels.pdf. The rest of the document is also worth reading.

Inserting an attenuator after the reference power splitter improves the directivity of the reference receiver by the attenuation value.

The Peregrine switch isolation is insuficcient for the transmission mode, Consider using two of them in series in each arm, a total of 3 switch ICs. Even for the reflection mode it is poor above 3GHz. Also you may want to use reflective switches, not absorptive ones to further improve isolation.

This is true and I made the same mistake myself. In fact I made it two times, because the receiver SP4T switch also had a poor isolation. The poor receiver isolation forced me to use 16-term calibration that can correct also for the leakage between the different receivers. You also need to use a calibration algorithm that can correct the isolation when doing a two port calibration. I think that SOLT+isolation is sufficient in your case since receiver to receiver isolation should be good enough. Try the 16-term calibration procedures in scikit-rf if all else fails.

[Bud]

It is better to address isolation issues at early stage and not rely on isolation calibration, which adds noise, prone to system drifts, adds complexity to the cal procedure and is not generally recommended unless making measurements close to the noise floor.

I'd also recommend to look at the filter banks isolation.

One more thing is I think i noticed no provisions for Board level RF shields, i think it was overlooked and boards have to be re-spinned with some space and copper islands added for soldering shields.

[henrikf]

Yes, while in theory leakage can be removed by the calibration in practice relying on it is not a good idea. Dynamic range of the receivers is limited and having to substract a large leakage term results in a worse accuracy. Drift is a much bigger concern when the error terms are large.

One consequence of the poor isolation I found out while writing the calibration code is that the switch terms can't be measured the usual way during the calibration. Because there is leakage through the switch, switch terms are not constant and depend on the DUT S-parameters. Switch correction must be done during the measurements by using the result of all four receivers. Result is less accurate since the signal on the reference receiver of the terminated port is very small and noisy. If the port isolation is good then during the normal operation it is enough to measure only three receivers as the terminated ports reference receiver result is not needed. I think there are some commercial VNAs that have four directional couplers, but only three receivers with reference receiver being switched depending on the active port.

[KE5FX]

What's up with the filter bank on your source?  Source harmonics don't generally hurt anything.  Yes, the leveling will work better without them, but you don't need leveling either with a four-receiver design.

[loxodes]

Thanks for all the comments and suggestions so far, I have a lot to keep in mind for the next hardware revision.

I've been sporadically working on the VNA and now have one port measurements working reasonably well out to 10 GHz (the upper limit of my synthesizers without adding in an external doubler). Other than replacing power splitters with used directional couplers off ebay, the hardware is largely unchanged from earlier posts. Attached is an updated block diagram of the current one port configuration.

Progress has also come from fixing errors in signal processing, calibration, reference clock distribution, chip initialization, power supply noise, bad solder joints in the signal path, and more..

Attached are measurements a mini-circuits VAT-3+ attenuator compared with Mini-Circuits published data, and of an ANNE-50L+ load and SF-SF50+ SMA barrel.

Also attached are here are un-calibrated measurements of my sdr-kits calibration kit along with the error terms from the calibration. I know these aren't ideal, I'm still investigating ripples..

My next steps will be to test 2 port measurements, and then to add in a frequency doubler for measurements out to (optimistically) 14 GHz. Once I've learned what I can from the current revision and have the software working reasonably well, I'll respin everything into a better shielded, less modular, and less expensive design.

Henrick Forsten recently updated his VNA with an interesting looking directional coupler (http://hforsten.com/improved-homemade-vna.html), I will try and replicate his efforts and replace my ebay-ed directional coupler modules with less-expensive-to-reproduce homegrown ones.

[loxodes]

What's up with the filter bank on your source?  Source harmonics don't generally hurt anything.  Yes, the leveling will work better without them, but you don't need leveling either with a four-receiver design.

Filtering is built into the synthesizer boards because I was originally planning to build a one receiver VNA without a reference path. I'll investigate removing/reducing the filtering on the hardware revision.

[cncjerry]

I'm glad I looked at your site before commenting on the directional couplers.  I bought Minicircuit DCs of all kinds and frequency ranges on eBay as new for dirt cheap.  I don't remember purchasing resistive couplers and my homebrew attempts didn't work out that well.  The best overall reflection bridge I bought came with a 0 to 500Mhz scaler analyzer I bought for like $150.  It has unbelievable response.  The transformers are made from what looks like RG-174 but it has just flat response up to several Gig.

i made a bunch of flat covers, CNC milled, for my N2PK VNA.  If you would like the same, send me a PM and I'll do them for the cost of shipping to compliment the nice work you did on the VNA.  I made them so they snugged right up to the board.

Jerry

[loxodes]

I'm testing a new revision of the network analyzer that takes into account some of the earlier comments and it is working reasonably well.
Two port measurements appear to work out to 13 GHz.

I think performance of the VNA is limited at higher frequencies by the lack of a proper cal standard. I'm using a $73 SMA cal kit (http://www.sdr-kits.net/Webshop/products.php?50, Mouser now also stocks the individual parts). If I can scrounge up access to a real VNA or calibration kit I could remeasure my kit and come up with a better model of the standards.

More detailed project updates are on a hackaday.io project page: https://hackaday.io/project/26213-vector-network-analyzer

[DaJMasta]

Very impressive work so far, even the change in the size of the nest of cables involved.

Better cal kits do sometimes show up on ebay, though I don't know how affordable they'll be if they're also in calibration... which may be desirable for this sort of application.  Is there a local test equipment rental place that could rent you a nice cal kit for a reasonable price?  If they do RF stuff, I'd expect at least pretty good ones to supply with rented out RF equipment, if you can get the calibrator alone for a little while, you could also get a little bit of characterization info on your kit.

[loxodes]

Better cal kits do sometimes show up on ebay, though I don't know how affordable they'll be if they're also in calibration... which may be desirable for this sort of application.  Is there a local test equipment rental place that could rent you a nice cal kit for a reasonable price?  If they do RF stuff, I'd expect at least pretty good ones to supply with rented out RF equipment, if you can get the calibrator alone for a little while, you could also get a little bit of characterization info on your kit.

I've been trolling ebay for 3.5 mm calibration and verification kits, but I'm hesitant to buy something used.. I may end building a TRL calibration kit or buying an (in)famous kirkby microwave type N kit and dealing with de-embedding type N to SMA adapters.

I didn't even think of test equipment rental places. I don't know of any within few thousand kilometers, but that could be a good option after I manage to relocate closer to civilization.