More results of my JREF project, please see Reply#243 at this topic. The results are somewhat mixed.
Results of analog compensation technique
1. A number of temperature-compensated JREFs were assembled and tested. The spacial separation of JFET and BJT (temperature sensor) induces a 0.8 ppm position dependency of the compensation and therefore also of the reference voltage. The output stability and noise is analysed by the calculating the Allan deviation of the reference voltage. Fig.1 shows the Allan deviation of a 10V reference with a 2N4391 JFET measured by a Keysight 34470A (after an 1 day warm-up) before final adjustment. The classical Allan deviation is green, AVAR is red, and MVAR is blue. A yellow slope in the double-logarithmic scale can be added to show the exponential law of deviation. The JRef noise is close to the K34470A noise at about 5s. The circuit also shows a linear drift (slope +1 at right end indicated by a disappearing distance of AVAR and MVAR) even while it is temperature-stabilised at the zero TC point. This drift is caused here by the very slow relaxation of the 2N4391 after a thermal 20 K step. Fig.2 shows the Allan deviation of a 2N4416 5V reference (Siliconix, 39 years old) in a non-stabilised room environment. The noise of a RF JFET is higher, but the complete circuit had +2 ppm drift within 2 months (measured by the K34470A, caution: specification +/-14 ppm in 2 months) and no significant hysteresis. The reference voltage is within 2 ppm about 5 minutes after power on, and within 0.4 ppm after 4 hours. Unfortunately, the JFET was selected not carefully enough: it shows some random telegraph noise (RTN) of 1 ppm, see Fig3.
2. The third and forth order compensation of the temperature dependency of a JREF reference by an analog circuit is possible, very stable, and it doesn't add significant noise. Within a 18°C to 35°C range, a temperature dependency of <0.2 ppm/K can be obtained. But, the adjustment of the analog compensation is difficult. In practice it is very complicated (if not impossible) to distinguish thermal hysteresis, drift of components such as resistors and trimmers, and circuit temperature gradients from the real thermal drift of the JFET reference voltage. The essential point is that modifications of hardware are required to adjust the compensation, but the thermal conditions are changed when you are performing these modifications. If you change a trimmer position, you have to wait about 1 hour before you can measure its effect precisely. Then you have to measure the reference voltage at >=5 temperatures. The available working time limits the number of different JFET types which can be tested. I have no realistic idea how to automise the adjustment. The adjustment is so much work that the analog adjustment technique is refused even while it is working well in a limited temperature range.
3. Several different JFETs were tested for low frequency noise by measuring the Allan deviation of a suitable reference circuit. No JFET was found which had only pure white and 1/f-noise. Even parts of the same type, manufacturer, and lot differ extremely with regard to low frequency noise like Random Telegraph noise (RTN). Concluding, all JFETs must be selected for minimum low frequency noise. According to my experience, <5% of JFETs are suitable as reference. Unfortunately, the RTN depends on temperature, working point, thermal history, and so on. A JFET selected for low noise might have high noise at a differing working point. You need luck. Even zener refs are not better in this point (potentially excluding the LTZ1000).
4. Some 4391 and 4392 JFETs of different manufacturers were analysed (Central Semi and dsi with lower noise, other m. were not available at this time). These JFETs had relatively low 1/f-noise superposed by a varying amount of RTN. However, the tested devices had a thermal hysteresis >5 ppm with 10 K steps and a relaxation time constant of many hours, which property prevents the application as reference. Please consider that 439x JFETs of other manufacturers might be better. The tested 2N4416 JFETs (original Motorola, 30 years old, and Siliconix, 39 years old) had a thermal hysteresis <1 ppm. RF-JFETs, which have separate pins for gate and case (4 pin JFETs as the 2N4416), must have any isolation layer between gate (mostly the JFET die) and case. The long-term stability of the Siliconix 2N4416 reference seems to be very good. It is an obvious guess that this layer reduces the thermally induced mechanical stress on the JFET die. It would be very interesting to know more about this subject.
Current status of the project
A new battery-operated JREF circuit will be tested soon (new PCBs are already delivered), which uses a digital compensation technique of the JFET reference voltage drift. The reference voltage is adjusted at several temperatures by generating an adjustment voltage. It is planned to use a third order interpolation parabola or spline between the temperature points. The JFET sees identical surrounding conditions at adjustment and normal operation except the surrounding temperature. It is interesting that the REF70 of Texas Instrument also uses a certain kind of digital compensation of the reference drift, however, it seems that the REF70 uses a completely different technique.