With so many discussions and data on precision voltage references, I've decided to resurrect my old (about 1993) idea of building my own voltage reference from scratch. At the time I had no means to measure it performance accurately, and even now, my two Keithley 2015 are struggling. The circuit I've built yesterday on my bench is deceptively simple, essentially one JFET and one resistor. I've used one half of U440 dual FET with a rather high (about 4V) cut-off voltage, connected a multi-turn W/W pot in place or R1 and adjusted the pot for a near-zero tempco of Vref output by heating the JFET with a soldering iron and looking at the voltage changes as it cools down back to room temperature. Then I've measured the resistance of the pot and replaced it with a couple of metal film resistors (standard type, 100ppm/C, nothing special). The beauty of this circuit is that the resistor changes reflect on the output voltage at about 1/100 of a value, i.e. 100ppm change in the resistance only creates about 1ppm change in the voltage - and the same applies to the supply voltage, about 1/90 ratio I've measured. I've used LH0070 as the supply regulator to avoid a possible supply influence though. The reference voltage for this particular sample of U440 is 3.75308V at 23C.
I did not "tune" the reference further and the circuit was just sitting on the bench overnight. I've recorded the reference voltage for over 12 hours and for about 5C temperature change the voltage changed from 3.75308V down to 3.75301 at the lowest point and came back to 3.75308 again after the room warmed up back to 23C (making the tempco about 4ppm/C, nicely linear, so it can be tuned better). I've tried to see a thermal hysteresis by first freezing the JFET with a freezer spray and then heating it with a soldering iron for about 5 sec, however the voltage recovered nicely back to 3.75308V after the device cooled down.
So here it is. If somebody with a better than mine measuring capabilities can have a go at this simple arrangement, I will be happy to see the results. I can see one serious problem with this circuit - the voltage reference point would be unique for each JFET sample , so it is difficult to use such a circuit in a production equipment. I will eventially build a proper boxed version with a scaling buffer to get, say, exactly 10V out of it. When I'll do it I'll post an update here.
Cheers
Alex