I have made a few revisions to improve the power supply – see schematic. First, the 15 volt zener at the bottom of the two MOSFET cascode shunt was deleted as unnecessary. Second, the 431 output filter capacitance was increased to 1000 μf to roll off its output noise above 0.8 Hz. Third, I now recommend running a 220 kilohm resistor from B- to each input tail current source, so I deleted the –C supply resistor chain. Fourth, I recommend floating both filament supplies rather than tying them to the high voltage supply, so I modified the voltage setting output resistors to all have the same value. Fifth I decreased the resistor chain values for the 431 chip the schematic shows the original values in parentheses. These are mostly deletions or changes in values but not in topology, i.e. if you leave out the component values the circuit looks pretty much the same as the original. Finally, I moved the capacitor from between reference and anode terminals of the 431 to across its cathode and reference terminals. Although this appears to be a very simple change, it significantly improved both stability and noise performance.
The 431 has an internal op amp with a gain of about 55 dB and a bandwidth extending to between 5 kHz to 50 kHz before rolling off. This op amp is not unity gain stable, which means it can and will oscillate. Despite this, the 431 chip is widely used as a voltage regulator in switch mode power supplies in all sorts of consumer electronics, such as PCs, laptop chargers, cell phone chargers, solar panel charge controllers, etc., so various techniques have been developed to stabilize it.
The regulator voltage is controlled by feeding back a portion of the output voltage into the reference terminal of the 431. Often a wide bandwidth is desirable to control both the DC voltage and cancel any AC variations across a broad spectrum. However, this design uses local feedback to the lower MOSFET via a capacitor to its gate to cancel any AC changes and noise, so we want to restrict the 431 to controlling the DC only. Broskie did this by rolling off the feedback signal using the capacitor between reference terminal and ground. But that did not affect the feedback loop gain, leaving lots of opportunities for oscillation. However, instead of limiting the feedback signal, if we limit the bandwidth of the 431 sufficiently, we both restrict its effect on AC voltage variations and stabilize it against oscillation.
This is done by removing the capacitor between the 431 reference and anode terminals, and instead connecting it between the 431 cathode and reference terminals. In combination with the resistor chain between B+ and the reference terminal, this forms a negative feedback compensation loop that rolls off the gain of the op amp at 6 db/octave above its corner frequency.
Since I modified the 431 resistor chain and output RC filter, I used trial and error to determine the needed feedback cap value. There were large oscillations with a 1 μf cap, while 3.3 μf produced smaller, slower oscillations but still managed to fry the 431. A 10 μf/50V polyester capacitor was stable with generally low noise, but random bursts of higher noise – however, the levels of the bursts were still below the 2 mV noise level in the original design. Increasing to 20 μf decreased the magnitude and frequency of the random noise bursts by about half.
With the original 4 megohm resistor chain for the 431, a 2.2 μf cap should give the same gain compensation curve, so the simplest modification is to remove it from its original position across the reference and anode terminals and connect it between reference and cathode terminals. On the PS PCB there isn’t room to mount a 2.2 μf cap next to the regulator chip but you can drill a couple holes next to the 200 ohm resistor on the opposite side from the 22 μf output cap and run insulated wires to the two outer leads of the 431 chip. Note that if you are using a polarized cap, the + terminal should be connected to the cathode and the - terminal to the reference. This is the opposite from the previous connection where the + terminal of the cap was connected to the reference terminal.
Replacing the TL431 with the quieter SPX431 dropped the noise voltage another 10 dB or so. With these changes, the broad-band power supply noise with a Fluke 189 DMM was decreased to around 0.16 mV into a resistor dummy load on the kitchen counter, with occasional bursts up 0.4-0.5 mV, so call it a -120 dB power supply. Not too shabby considering its simplicity. DC voltage overshoots by a volt or so but then remains within 0.1 volts of its target voltage due to the excellent temperature compensation in the 431 chip.