I saw this recently:
One thing I like about it is that it drives the LO port of the mixer with square waves which should work better (more linear output amplitude vs. input amplitude and more stable gain with temperature) than the sine waves used in the N2PK VNA. Also, it is better than the DG8SAQ VNWA3 in that it does not rely on using non-harmonic spurs of a DDS as the LO for higher frequencies, where the small amplitude of the spurs might lead to mixer conversion gain that varies with temperature etc.
I think it could be improved quite a bit by not just having one receiver, for a couple of reasons. One reason is that the leakage from one channel to another across the switch on the receiver is likely to be maybe -50dB so that it would be hard to calibrate it out well enough to measure insertion loss of e.g. 100dB. Another reason is that when the receiver is switched to measure a different signal, the impedance seen by the coupled ports of the directional couplers changes, so that a different amount of signal gets reflected back into the couplers, which would change the result that is being measured, in a way that I don't think can be fixed by calibration.
Once there are more receivers, to allow high insertion loss to be measured it would be good to have an ADC with more than 12 bit resolution, or at least a gain switch to allow a greater range of signals to be measured. I was thinking of using the AD7608, as this could sample both mixers of each port of a 4 port network analyser simultaneously, and has built-in anti-aliasing filters which should match well as they are on the same chip. An AD8253 between each mixer and the ADC would allow gerater dynamic range in spite of the limited channel-to-channel isolation of that ADC, but would add a fair bit of cost. The IF would probably need to be in the range below 10kHz instead of 2MHz but I don't see a disadvantage in that, and it might make IF leakage across the PCB easier to control. I'll have to think about whether the LO and source phase noise matter in choosing the IF.
If the directional couplers are replaced with bridges built using the differential input of the mixers (LT5560) to perform a subtraction operation at RF, then the frequency range could be extended down almost to DC, and the PCB area of the couplers could be avoided.
To drive all of the mixer LO ports, I was thinking of using the ADCLK948 or similar, ideally with DC-coupling to the LT5560 LO inputs, if the DC levels can be shifted to be correct (perhaps by using weird supply rail voltages). The LO signal does not need to go to a frequency as high as the maximum source frequency of the VNA, because the mixers should be quite sensitive to the 3rd and 5th harmonics of the LO signal as well as the fundamental. One input of the ADCLK948 clock buffer could be driven directly from one output of an ADF5355 synth. The other output of the ADF5355 would clock a differential D-flip-flop (e.g. NB4L52) that has its Q-output driving the other input of the ADCLK948, and the D-input of the flip-flop would be controlled by a FPGA that also receives the clock signal of the D-flipflop, so that lower frequencies than available from the ADF5333 (or arbitrary bitstreams) could be sent to the LO ports of the mixers. Depending on the isolation between the RF and LO ports of the mixers, and depending on the output-to-output isolation of the ADCLK948, it is possible that an extra stage of LO buffering for the two mixers of each port might be useful to eliminate "mixer bounce", but that would be expensive in parts count and cost.
The filter bank used in the VNA source seems like a good idea, but for the lowest frequencies, the source PLL could be used as the clock of a DDS. If the DDS is programmed with tuning words that are always powers of 2, then it should not generate any non-harmonic spurs at all. In this way, it could be used as a divider to generate low frequencies in a way that does not require much filtering. For higher frequencies, the output of the PLL chip could be put through a filter bank as usual.
If the RF PCB can be made small then it might be worth putting heater resistors and thermistors to ovenise the whole PCB, so that its error terms are more stable between when it is calibrated and when it is used.