Hi,
I'll try to answer the questions. My first approach was to build a series of noise sources of high-level characteristics, calibrated, with frequency 1-500MHz and costing less than 200 Euros.
Looking at the solutions on the market I saw that, apart from the HP sources which are similar if not the same as the Noise com, it consisted of a capacitor, a diode and a bias resistor in addition to the input decoupling. consequently I built a test set by mounting a spider web to test various diodes. I was lucky that a friend lent me a Noise Source analyzer and an HP source to do the first tests. The first diode was a 1N4148 which works very well with a high noise level which however was not flat, the ENR variation at low frequencies was more than 10dB higher than at 500MHz which is also its maximum operating frequency.
So I thought of using a classic BFR93 reverse biasing the BE junction which works at higher frequencies but it too had a diode-like linearity at low frequencies, so to linearize I decreased the value of the output capacitor from 1kpF to 300pF. A solution that works but I don't like it because in this way part of the stability of the level, below 100MHz, is shifted to the temperature stability of the capacitor.
So I bought a dozen of BFR90, BFR90A and BFR91. These BJTs have very good frequency linearity up to 2 or 3 GHz and beyond. At this point I understood how the diode bias current works and I mounted a constant current generator which must also have the characteristic of having a fast response to follow the 28V which are switched on and off at a frequency of about 1kHz. In the end I opted for a different circuit from that of the heads of the EME conference in Florence, much simpler but functional. This is composed of a FET and a resistor which determine the BIAS current which is in my case less than 2mA.
I would like to add that at the output of all the sources I have mounted a 50Ohm 10dB attenuator to have a low SWR, as the diode and the generator components do not have an impedance close to 50Ohm, especially when the diode is in the off state. It is a configuration that all manufacturers use. In my case the worst return loss is 24dB like the HP version I have.
But back to the bias current, I've noticed that it has a different effect depending on the frequency. The rule is that for frequencies lower than 1-2GHz the variation of the current causes the noise level to vary greatly and much less at higher frequencies. So I took the noise level at 1.5GHz as the base noise level of the single diode and I varied the bias so that the response curve aligns with that value even below that frequency.
In addition to this I have verified that even using the same transistors, but from different lots, the generated noise level varies. Don't ask me for scientific explanations.
For example, two BFR90s from different batches generate, one 13 and the other 16db (after 10dB of attenuator). Therefore, the rule is that these devices are all quite good but they must be calibrated individually, each one has its own specific current to have excellent linearity.
Now let's talk about pcb, I decided to use a material for microwaves, the FR4 loses about 0.2dB at 2GHz, not much, but I preferred to start from an excellent quality support also because I can use it in the future for higher frequency sources.
I did the calibration with the Agilent NFA which gives me the response with a graph on a color LCD display. Having a sample source, first I normalized the trace, then I connected each of my sources and the reading difference with the markers will be subtracted or added to the ENR table of the reference in order to have its own personal table.
It is very difficult to do bias adjustment with Noise Figure Meter without graphic display. I hope I have answered your questions.
additinal note: testing several diodes I found one is +/- 0.3dB from 10M to 3GHz.
Luciano