Having six LTZ1000 in the drawer i studied a little how to use them as an array in an external oven. From my ADR1399 experiments i learned how to make ovens big enough for large filter caps and for 7 to 10 V gain stages.
A known way to adjust the TC of a LTZ1000 reference is a small resistor in series with the zener. It forms a voltage divider together with the 120 Ohm resistor that determines zener current. In effect we amplify the compensation of the temperature dependent Ube. The two resistors need to be precision parts and we can tune the gain a little using reasonable quality tuning resistors. To reduce number of iterations one needs to adopt a numerical model.
Using parts from the drawer i made two cells of the array and adjusted the first reference to zero TC using a hairdryer on the bench. The adjustment resistor of 18,33 Ohm consists of 2x UPW25 10R + MF 1% 220R. The second LTZ1000 got the same compensation and the board went into the incubator for temperature sweeps.
In the diagram readings are shown as deviation from the average voltage during the sweep:
Ref1 Average = 7.186412 V
Ref2 Average = 7.165433 V
The result shows a zero TC of the first reference at about 35 °C, a nice result. The real chip temperature will be a little higher, but not much (non-A version, long legs, 4 mA * 6.6 V = 26 mW).
The second reference was overcompensated. After some iterations of fine tuning it arrived at a compensation resistor of 16.91 Ohm in order to have the same zero TC temperature as the first reference. The agreement between both references is about 0.07 uV/K or 0.01 ppm/K and stable.
From the tuning process i determined the TC of Ref2 Ube as -2,07 mV/K. Ube is 488 mV at 20 °C.
Both reference TC measurements exhibit a very similar curvature. Without the tuning resistor the zero TC temperature would be higher, maybe closer to the recommended LTZ1000 oven temperature.
Soon more precision resistors will arrive in order to complete the other four cells of the array. I hope the observed 1 ppm "hysteresis" will then disappear.
Regards, Dieter
PS:
Measurements are from a Keithley 2700 with 7706 multiplexer plugin and of remarkable quality for this so-called 6.5 digit meter. Each data point is the average of 30 samples of 5 PLC ("slow"). The calculated difference between the two LTZ1000 references has a standard deviation of 0.91 uV over the whole sweep.