Linear LT1236LS8 may be an alternative to MAX6226 if MAX6226_50 cannot be made to work reliably.
EEVBlog thread on LT1236LS8 internals
Both use ceramic LS8 package with copper pads underneath the device. This gives improved stability, particularly to moisture and PCB flex, but makes soldering difficult without a reflow oven. Hot air tends to blow the lightweight ceramic chip off the board, so holding in place with tweezers is required.
Vref board needs to provide both +5V and -5V to drive the pitch DACs.
Good initial accuracy ±0.05% (±2.5mV) for A grade; although less good than MAX6226_50 ±0.02% (±1mV). But it is trimmable, using a resistor, diode and multiturn cermet trimmer.
Pin 5 is trim, unconnected if not used; different from MAX6226_50 where pin 5 is output sense.
Fairly good at 2ppm/C typ, 5ppm/C max, not as good as MAX6226_50 1ppm/C typ, 3ppm/C max.
Bourns wirewound trimmers with 50ppm/C are fearsomely expensive (3057Y-1-503 is $32.91/1) compared to the 100ppm/C cermet types (3296P-1-503LF $2.42/1). As this application is ratiometric, the tempco effect should be diminished.
See Bourns® Precision Trimming Potentiometers with Improved Linearity and Lower CRV
Very good at 3ppm.
The data sheets of many references specify long-term drift—typically about 25 ppm/1000 hr. This error is proportional to the square root of elapsed time, so 25 ppm/1000 hr ≈ 75 ppm/year. The actual rate is likely (but not certain) to be somewhat better than this as the ageing rate often diminishes after the first few thousand hours. So, again, we have a figure around 14 bits. Choosing Voltage References
20ppm/1k hr (100μV), less good than MAX6226_50 11ppm/1k hr (55μV). But trimmable after pre-ageing.
Load regulation very good: ±25μV/mA max, and the load impedance is high and constant.
This design drives two DACs, so the lack of a kelvin connections is less of an issue.
The DAC boards provide their own force/sense buffers to drive the voltage reference inputs on each DAC.
Good at 4 (typ) to 12 max μV/V for supply voltage between 7.2V and 10V.
So a well-regulated supply is helpful.
Use same regulated +9.5V supply as pitch DACs. Off-board, to protect from thermal effects.
Spectral noise is 3 μV p-p 0.1-10Hz Flicker noise is 3.5 μV RMS (typ) at 10Hz-1kHz.
These specs are not directly comparable to MAX6226_50 due to different units and bandwidths.
A non-auto-zero, e-trim op-amp such as OPA2192 is suitable, and input bias current is even lower.
For the reference buffers, the AD8676 dual amplifier is recommended, based on its low noise, low offset error, low offset error drift, and low input bias currents.
Egan, M. The 20-Bit DAC Is the Easiest Part of a 1-ppm-Accurate Precision Voltage Source
| noise | offset error | offset error drift | input bias current | |
|---|---|---|---|---|
| AD8676B | 125nV RMS | ±12/50 µV | ±0.2/±0.6 μV/°C | ±4.5nA |
| OPA2186 | 0.1μV (100nV) p-p | ±1/±10 µV | ±0.01/±0.04 μV/°C | ±4.8nA |
| OPA2192 | 4 μV p-p | ±8/±50 μV | ±0.1/±0.5 μV/°C | ±5/±20 pA |
So the OPA2186 is comparable on some criteria and much superior on offset and offset drift. OPA2192 worse on offsets (still very good though) and better on input bias current.
The input bias current specification of the reference buffers is important, as excessive bias currents will degrade the dc linearity. The degradation of integral nonlinearity, in ppm, as a function of input bias current, is typically:
Extra INL error = (0.2 * Ibias) / (Vrefp - Vrefn)^2
where IBIAS is in nA; VREFP and VREFN are in volts. For example, with a ±10-V reference input span, an input bias current of 100 nA will increase the INL by 0.05 ppm.
Egan, M. The 20-Bit DAC Is the Easiest Part of a 1-ppm-Accurate Precision Voltage Source
Maybe try a non-autozero non-chopper op-amp. OPA2192 looks suitable: Vos ±8μV typ, ±50μV max, drift ±0.1 typ, ±0.5 max μV/°C.
See figs 32 and 33 in the datasheet:
See Figure 69. Precision Reference Buffer in the datasheet (driving a 10μF capacitive load with great stability).
Pair of close-tolerance low thermal drift resistors for the VrefN op-amp. Use of Vishay matched-pair resistors suggested on EEVBlog such as these 10k/10k 0.01% ratio, 25ppm/C. Is it feasible to make a board that accepts either?
Decoupling caps.
Use small serial resistors for the larger decoupling caps because too-low ESR of the ceramic caps might affect stability. Can test with either 5R resistors or solder bridge to see the effect.
Power input: 4 wires +9.5, 0V, 0V, -9.5.
1μF and 100nF input caps, 1μF (or less; that is a maximum) and 100nF output caps, all C0G ceramic. Later went for X7R on the 1μF for cost and footprint reasons.
Because of internal current drain on the output, actual worst-case occurs at ILOAD = 0. Significantly better load transient response is obtained by moving slightly away from these points. See Load Transient Response curves for details (Datasheet, "Capacitive Loading and Transient Response")
Thus, 10k load resistor draws constant 0.5mA from reference (assuming drain from the op-amp input currents is negligible in comparison).
Voltage output: pair of 4-pin headers
- +5 +5 0V 0V
- -5 -5 0V 0V
Full bottom ground plane; top ground connections with vias by caps and GND connection, runs under device
Pin 1 of LT1236 is indicated by a circular mark and is aligned to the top-left corner on this board (indicated by small dot on silkscreen).
Pin 1 of OPA2186 is on the side with the bevel, and is aligned closest to R5 on this board.
Input cap has 5R serial resistance to avoid ringing. Electrolytics were better here than ceramics.
In critical applications, a 10μF solid tantalum capacitor with several ohms in series provides optimum output bypass. (Datasheet, "Capacitive Loading and Transient Response")
2 layer board of 1.57 x 1.03 inches (40.0 x 26.2 mm) ordered 2024-02-10
top and 
bottom
(1) LT1236AILS8-5#PBF Ceramic LCC $11.77/1 = $11.77 GOT 4 24 Feb 2024 IC1
(1) TI OPA2192IDR SOIC-8 $4.63/10 = $46.30 GOT 10 Dec 2023 U1
(4) Kemet C1206C104K3GEC7210 25V 100nF C0G 1206 ceramics $0.051/100 = $5.10 GOT
(2) Kemet C1206X105K3RACTU 25V 1μF X7R 10% 1206 ceramic $0.839/10 = $8.39 GOT
or
(2) muRata GRM31C5C1E474JE01L 25V 0.47μF G0G 5% 1206 ceramic $0.474/10 = $4.74 GOT 10 24 Feb 2024
(2) Susumu RG2012V-182-P-T1 1.8k 0.02% 5ppm 0805 $2.30/10 = $23.00 GOT 6 ordered 8 Dec 2023 GOT 6 more 24 Feb 2024
OR
(2) Vishay Precision Group Y16292K50000T9R 2.5k 0.01% ± 0.2 ppm/°C 0805 $12.74/1 = $25.48 at minimum (need more for multiple boards, for testing).
- (2) Yaego RC0805JR-075R1L 5.1R 5% 0805 thick film $0.026/10 = $0.26 GOT 10 24 Feb 2024
- (1) Yaego RC0805FR-0710RL 10R 0805 $0.027/10 = $0.27 GOT 10 24 Feb 2024
- (1) Yaego RC0805FR-0727KL 27k 0805 $0.016/10 = $0.16 GOT 10 24 Feb 2024
- (1) Yaego RC0805FR-0710KL 10k 0805 $0.016/10 = $0.16 GOT 10 24 Feb 2024
(1) 1N4148 Diodes Incorporated 1N4148W-7-F SOD-123 $0.095/10 = $0.95 GOT 10 24 Feb 2024
(1) Bourns 3296P-1-503LF $2.13/10 or $2.42/1 get 5 = $12.10 GOT 5 24 Feb 2024
Ordered OSH Park 10 Feb 2024.
Right-angle 0.1" connectors (1) Adafruit Break-away 0.1" 36-pin strip right-angle male header (10 pack) $5.95 https://www.adafruit.com/product/1540 out of stock
TE 9-103329-0 40-position $2.31/1 GOT 3
Double ended connectors Adafruit Extra-long break-away 0.1" 16-pin strip male header (5 pieces) $3.00 GOT https://www.adafruit.com/product/400
pins are (measured with micrometer) 45.06 ÷ 2 = 22.53mm apart
- Design v0.1 Vref board with dual op-amp
- Fab v0.1 board
- Check existing component inventory
- order Vref, op-amp (same as pitch DAC uses)
- Build, test
- Burn-in
- Do stability analysis for innie OPA2192 with 2 R and 1 C
- Design and fab temporary power board to test for ripple, noise, need for isolation resistors