Building a 22 GHz network analyzer for under $1000

[EggertEnjoyer123]

I am thinking about trying to build a DIY network analyzer based on the LiteVNA architecture.

There are several issues (more like tradeoffs to make the device cheaper) with the LiteVNA design. They are as follows:
1. Bad directivity with the bridge. The bridge starts to lose directivity at around 6 GHz, and the noise floor for S11 quickly increases to about -15dB or so at 8-9 GHz. To be fair, it was never designed to operate that high.
2. Bad mixer. For cost reasons, the mixer used is the AD8342, which is only supposed to work up to 3.8GHz. This is why the noise floor increases quickly for S21 measurements after 4 GHz.
3. Bad isolation with the RF switches. An RF switch is used to switch between the reference, reflected, and transmitted signals. The RF switch (MXD8641) is rated only to 3 GHz. That's probably another reason why the noise floor increases past 3 GHz, at least with S21.

My idea is to use easily obtainable used RF modules which can be found at ham swap meets. I have a couple 7-18 GHz high directivity couplers from Narda, which have >25dB directivity. I also have several wideband mixers which cover the entire frequency range. My plan is to use two LMX2820 chips, which can be had for about $100, and a few RF amplifier ICs (like the MAAM-011100, which costs $12) to power the mixers and generate the signal for Port 1. The two chips will be set to have an offset of 10kHz, allowing a cheap audio ADC to be used to digitize the signals. I could also use mechanical RF relays to solve the isolation issue. Any issues with my idea?

[joeqsmith]

How low of a frequency do you plan to support?  Are you planning to support more complex calibration methods?   Did you try your 7-18GHz couplers are 22GHz and found good enough performance?   

Keep us posted.

[Gerhard_dk4xp]

Have you seen this here: 
<     https://www.analog.com/en/products/adl5960.html     >     ?

regards, Gerhard

(who is just writing a driver for LMX2594/95 )

In the end, directivity  performance is the result of a lot of math
invested in 12-term-error correction etc.

[szoftveres]

If I could change only one thing with the existing LiteVNA, I would probably modify it for lower output power (e.g. attenuate the source by 20dB, and add 20dB LNAs before the mixers). It would enable measurement of active devices in their linear region without giving up of dynamic range of the VNA.
Beyond that, an option to be able to switch between "normal" or "-20dB" output would be the ultimate usability improvement for me.

[EggertEnjoyer123]

I bought a LMX2820 test board from Aliexpress for around $120 (Link: https://www.aliexpress.us/item/3256806758049656.html, it was also on sale when I bought it).

The chip seems to work well, but amplitude falls off quickly after 20 GHz. I'm not sure if this is due to some random resonance in my setup or not (probably need to tighten down all the connectors), but the amplitude does drop sharply after around 20 GHz. I was able to get 0dBm out at 20 GHz but only around -5 to -6dBm out at 22 GHz. At lower frequencies the output amplitude is close to 7dBm.

I think a lot of it is due to FR4 losses. In any case -6dBm + 12 dB from an amplifier is enough to drive a mixer (but there probably needs to be some kind of leveling to avoid damaging the mixer at lower frequencies).

[ftg]

Asking FR-4 and SMA's to do 20GHz might be a bit much.
I'd blame the lossy FR-4 and then the PCB->SMA transition might not be the best available.
That said, the manufacturer does claim that the SMA's used are infact 3.5mm connectors made for 26.5GHz.

The claimed output power is 0dBm on 20GHz, -0.8dBm on 21GHz, -3.8dBm on 22GHz and -3.5dBm on 22.5GHz.

I also grabbed the images of the from that Aliexpress listing so that they can be seen more easily and will be available even when that listing is gone.

[EggertEnjoyer123]

Here's my annotated version of the board.

There are some dubious design choices (as with all cheap Chinese boards). The main ones are:

1) Using the same loop filter components as the datasheet while using a different PFD frequency (though 140 MHz vs 200 MHz might be close enough that this doesn't make a difference, assuming they are multiplying the input frequency by 2 and then 7)

2) Loading the output of a (presumably CMOS) oscillator with 25 ohms (2 x 50 ohms). I'll probably try removing them to see if that does anything.

3) Not using C0G capacitors for the loop filter

They're also using LC filters for the power rails, and I doubt they've considered resonance.

Anyways, it looks like I might be able to get the chips for $50 each off LCSC, which is much better than the $107 that Arrow charges (or $123 for Mouser). I've had good luck buying parts off of there, and they should be legit parts.

[jjoonathan]

"What is the bandwidth of your loop filter?"

"About 10 degrees Celsius"

[EggertEnjoyer123]

Removing both 50 ohm resistors seems to have decreased phase noise by a few dB (which makes sense, since loading a CMOS oscillator with 25 ohms is not a good idea).

Other changes I tried (such as more decoupling and adding resistors to get rid of potential resonances) didn't do much at all.

The power supply filtering is also not perfect, since some switching noise gets in. When I power it off a crappy USB power bank, you can see 500kHz noise from the switching.

Also, it appears that the PFD frequency is only 10 MHz, which is different from the datasheet's 200 MHz. This might mean that the loop filter needs to be changed. I determined the PFD frequency by looking for reference spurs, and there seems to be one at 10 MHz no matter what frequency I choose.

Edit: Dropping the PFD frequency by 20 times means the loop filter's bandwidth is now 20 times less - which is not very good.

Also, here is the expected phase noise (with 200MHz PFD frequency) and here is the actual phase noise (with 10 MHz PFD). You can see that the bump around 30kHz lines up with my spectrum analyzer measurements

[EggertEnjoyer123]

I'm currently working on the PCB design for the new VCO.

I found two couplers from (presumably) some HP VNAs in my rusty parts bin. The first one had 12dB directivity (and a bent APC-7 connector which I "fixed" using pliers and quite a bit of force), while the other had a bit over than 20dB (I measured from 0 - 9.3GHz, which is the best my LiteVNA could do). Before someone gets mad at me for "fixing" RF connectors, I connected a piece of coax to act as a "port saver", and then I marked all the bad connectors with Sharpie just to make sure I never use those connectors again with good parts

Here's what's inside the 12dB directivity one. It is much simpler than I expected, and a lot wider in bandwidth (2GHz to >9.3GHz) than ones made with microstrip. There is a triangular piece of metal which is clamped with pressure. Moving the triangle with a screwdriver didn't do much in terms of return loss or directivity, as measured by my cheap LiteVNA. The directivity was always around 12dB at lower frequencies.

[EggertEnjoyer123]

I have finished the schematics for the boards.

The MCU board is going to be 100x100mm and 2 layers, while the two smaller 50x50mm oscillator boards will be done using 4 layers. The MCU board consists of the STM32F407, which should be able to handle I2S, as well as a TLV320AIC3204 ADC, which will sample the downconverted signals from the mixer. I am planning on using an IF frequency of 12kHz, which is exactly a quarter of the sampling rate of the ADC (which will allow me to do math very easily as sine/cosine is always 1, 0, or -1). The MCU board also contains the power supply for the RF relays, as well as some drivers for them. Finally, the MCU board also controls the gain of all the amplifiers, allowing some crude leveling to be done. I have some more directional couplers, so it might be possible to do ALC later with cheap diode detectors.

The oscillator board contains the LMX2820, as well as protection circuitry for the RF amplifiers (MAAM-011100). The comparator should make sure that the transistor can't have 5V connected to it when the negative bias for the gate is not present.

I've also ordered all the parts already. The MAAM-011100 was bought on Richardson RFPD (https://shop.richardsonrfpd.com/Products/Product/MAAM-011100) for $5.2 each, while the LMX2820 was bought on LCSC for $49.7 (https://www.lcsc.com/product-detail/RF-Misc-ICs-and-Modules_Texas-Instruments-LMX2820RTCT_C2871859.html, you will have to submit an RFQ). All the other parts cost basically nothing.

[EggertEnjoyer123]

I've ordered the boards (after making some changes to my PCB schematic and layout). If anyone wants to mess around with the design files I also shared them (use JLC04161H-7628 Stackup for the 4 layer board).

I also was finally able to get my hands on a transfer switch and an absorptive relay for $5 each at the ham swap meet. This will simplify my design a lot, as I no longer need 3 reflective relays to switch between the transmitted and reflected signal. They are both DC-18GHz (one of them has the markings rubbed off but the part number seems to end in 33S60). I guess I might have to change my driver circuitry since the relays take TTL inputs instead of 28V (the controller is inside the relay).
 

[points2]

hi EggertEnjoyer123,
1st, thanks for sharing about such a high-profile project !
far from what I could... 

2nd,
do you have an idea about / can you rank : what matters in this "upgrade from a nanoVNA" so to speak...?
To jump to 22Ghz, of course you can reply : all matter ! and you're right.
but,
what's the influence of the oscillator
- vs the perf of a such device ? (nano-VNA & your project) ?
- vs the other parts to improve : PCB design / SMA in/out / perf of the IC.. etc...

my e-tweak level is currently about oscilllators

[EggertEnjoyer123]

hi EggertEnjoyer123,
1st, thanks for sharing about such a high-profile project !
far from what I could... 

2nd,
do you have an idea about / can you rank : what matters in this "upgrade from a nanoVNA" so to speak...?
To jump to 22Ghz, of course you can reply : all matter ! and you're right.
but,
what's the influence of the oscillator
- vs the perf of a such device ? (nano-VNA & your project) ?
- vs the other parts to improve : PCB design / SMA in/out / perf of the IC.. etc...

my e-tweak level is currently about oscilllators

The VNA works by sending out a signal of a certain frequency and measuring the reflected/transmitted signal. The LiteVNA has a PLL IC (MAX2870) that generates the 6.3 GHz signal (I blelieve they are cheating and using the chip out of the rated frequency limit, but it works anyways). You can go past 6.3GHz, but then you're using the harmonics of the PLL IC to generate the signal (so at 7 GHz, the oscillator IC will actually send out 3.5GHz, and we measure the 7GHz harmonic). The downside of this approach is that with amplifiers, it could be the case that the lower harmonic saturates the amplifier, which would throw off your measurements). Of course, my design isn't much better (since the LMX2820 outputs a square wave at lower frequencies), and commercial VNAs have filter banks to deal with this issue (but that would cost too much for me). For a VNA, you actually need two oscillators (one to drive the mixer and one to drive the output of Port 1). The second oscillator is set to a slightly lower frequency than the first oscillator (for me, probably somewhere around 10-20kHz lower, since the output would then be in the audio range, and I can use a cheap audio ADC to digitize it). That way, after mixing you get a very low frequency which is much easier to analyze and get the magnitude/phase information from.

The architecture of my design is going to be pretty much the exact same as the LiteVNA, but with higher quality parts. You'd actually be surprised by how much the LiteVNA creators got away with to save money. The mixer that is used in the LiteVNA is only rated to 3.8 GHz, which is why the S21 noise floor degrades after that frequency. The switches in the LiteVNA are actually only rated for 3 GHz, and leakage becomes a big issue at 6 GHz (another reason why the noise floor increases). The directional bridge in the LiteVNA is built using ferrite baluns, and those also lose directivity at high frequencies (causing your S11 noise floor to increase). In my design, I'll use mixer modules that are rated to 26.5 GHz, some 18GHz directional coupler modules (hopefully they still work at 22 GHz), and mechanical RF switches to solve all of those issues. The LiteVNA also is unable to switch the ports (you can get S11 and S21, but not S12 and S22), but my design will have a transfer switch.

PCB design is actually a lot easier than you'd think. Since all the chips are matched, routing the PCB is just making sure all the traces are 50 ohms impedance. The SMA connector just has to have a center pin with the same thickness as the trace. The return loss of the entire system is dominated by the amplifier MMIC I'm using, which could have S22 as bad as -5dB. The MMIC return loss will definitely dominate over everything else in the system. To fix the return loss issue I'll probably just put an attenuator after everything. Also, I'm still using FR-4 because the Aliexpress board uses FR-4 and the output amplitude is still high enough to get to the 10dBm needed to drive the mixers (if I use two amplifiers in the middle). Switching to a good substrate like Rogers would cost way more than the $5 I spent on the extra amplifier. The only other real gotcha is making sure all the SMAs are torqued correctly. Unfortunately, the only torque wrench I have is from Aliexpress.

I also got all the boards back, and I'm working on assembling everything. So far I found a mistake with the TL431 on one of my boards . Going to have to bodge that one...

[points2]

Thanks EggertEnjoyer123 for this precise feedback.

To be honest, I had a look at a NanoVNA-H4 a while ago, noticed a few things to test in order to improve it, but I switched to another topic in the mean time.

A full PCB design is out of my current skill (but it's on my road map :-) ).
I just do some works looking at low level PCB behavior, I mean : PDN.

Given the PDN influence on any transient state (clock signal is a "basic" transient state that occurs on a PCB at low level), a specific care upon the PDN of the main crystals/oscillators can improve the overall efficiency of the device (tested on a FMC, fiber-media-converter vs RT applications).
So, why not in a such a device with transients all over the place ?

Oups ! EggertEnjoyer123, here is just my humble input... and your feedback is appreciated (bad or bad :-) )

[Bud]

Why do you need 22GHz?
Do you have an idea how are you going to calibrate such device in-place?

[EggertEnjoyer123]

Why do you need 22GHz?
Do you have an idea how are you going to calibrate such device in-place?

Mostly just for fun (though it would be nice for 10 GHz stuff if I ever get into doing that)

I'd probably need to buy an actual calibration kit (which would cost several times as much as my entire project) if I ever wanted to get good results. But for now I could probably just see how far I can get with the LiteVNA cal kit.

[EggertEnjoyer123]

I got the boards back and I assembled some of them. So far there's been a few minor problems (I forgot the ferrite bead for the crystal oscillator, and I messed up the pinout of one of the TL431s).

Programming the firmware is quite annoying, since bugs will cause the chip to sometimes work and sometimes not work. In the first spectrum analyzer picture, we have the amplifier running at a lower control voltage (-1.5V vs 1V), and the amplitude is about 25dB less than the maximum output. I also had the attenuator in place (made of 0201 resistors). Each peak is spaced apart by 1GHz. There is definitely enough here to drive a double balanced mixer at full power. Unfortunately the firmware broke right after I took the picture, and it refuses to lock to 20GHz (but it locked to 22 GHz). The spurs were mostly gone after I put in a ferrite bead to kill any 100MHz noise from the crystals from coming back through the power rails. It looks like I might need more filtering in a later revision.

I ended up removing the attenuator to get more power, since I found that there wasn't enough output power at 1V control voltage. At -1.5V control voltage, the amplitude increased by 4dB, but the amplitude only increased by 1.4 dB at 1V control voltage, which means my amplifiers are in saturation. I measured with a power meter and I got 8.6dBm, compared to the spectrum analyzer's 5.6dBm. Assuming that there's 4dB loss through all the connectors and through the FR4 (which is very lossy at 22 GHz), I think 8dBm would make sense since the datasheet has 12dBm for the P1dB at 20 GHz.

I'll probably have to fix all the bugs this weekend and maybe try to get more power by increasing the gate voltage, which should allow more current to flow through the device.