Thanks for sharing, it looks good!
The self-heating will depend on the sensor, of course. Also, since thermistors have such a non-linear response over the temperature range, all temperature performance has to be calculated across the range. (I'm sure you're already aware.)
Target sensor is GA10K3A1IB TT/Measurement Specialties with a dissipation factor of 0.85mW/K.
There are much cheaper ones with good enough stability, so option to cut cost -
NTC long-term drift measurement p.58ff.
The sensor will be calibrated anyway, but you are right, with rising temperatures the required precision rises - especially in this not ideal configuration with high reference resistor (tradeoff for lower self heating).
I suppose you don't connect the NTCs with long cables, given that it's a 2-wire connection? Your graphs say the resistance is 40 kΩ at -4 C and 1 kΩ at 87 C. Not too difficult resistances to work with, but without running the numbers I guess the tempco of the leads still is important at high resolution.
For my primary purpose max lead length will be 1/2m.
NTC has ~2.5kΩ@60°C with ~90Ω/K, e.g. a 30AWG wire has around +-41mΩ/m in +-30K span.
This translates to absolute max +-1mK due to tempco of 1/2m wire - so even longer and smaller diameter wire is not an issue.
Do you use tantalum capacitors as the 3D hints? If so, why?
The C1, C2 caps are marked "MKS" and look orange in 3D. Looks like these: https://www.wima.de/en/our-product-range/metallized-capacitors/mks-4/
Do they have any advantage over regular MLCC?
Yes, yellow cases are tantalum, but could be replaced with MLCC -
followed reference design which uses tantalum (p.20).
Kleinstein already gave the answers, in addition the cheap MLCCs (X, Y, Z) are microphonic.
R4 and C2 are the anti-aliasing filter that "attenuates the frequency content around DMCLK and keeps the desired accuracy over the baseband of the converter" (quoting the MCP3911 datasheet p 35). Is your placement of this filter before the buffer better than after? Can't the buffer introduce new noise?
Is R6/C17 a low-pass filter for the buffer? What's such a filter called?
What's the purpose of D1, D2? Won't they short out the signal?
C1/C2 MKS 1µF are for anti aliasing and reducing noise (emi) in conjunction with NTC & R1/R2, but are not effective at line frequency or VHF.
1µF increases response time of sensor only by some ms, which is insignificant as sensor has ~1s response time.
Would be worth checking if an anti aliasing filter between opamp and ADC input is better - particularly if a more noisy opamp is used.
R3/R4 give some input protection.
R5/C16 & R6/C17 in FB: R for input protection, C for compensation of input/stray capacitance to GND (keep gain close to one at high frequencies).
Most opamps have input protection clamping diodes between inputs and inputs to rails, which can handle only a couple of mA continuous.
D1&D2 are low current jelly bean diodes (e.g. BAS316) for input protection of ADC, they are only slightly biased when measuring low temperatures @high PCB temperature.
thanks in advance for your contribution.. I'm curious, which precision NTC you have used, and which reference thermometer, and how you did the physical calibration and characterization.
Target sensor is GA10K3A1IB TT/Measurement Specialties, but all (released) measurements were made with dummy resistors (PTF56 5ppm/K @21+-1°C).
Calibration will be against a SHT35 in thermal chamber, which were used for TC characterization already.
Have 3xSHT35, which were calibrated against each other, ice-bath and boiling point of water and were in accordance to each other mostly within noise (+-10mK), ~+20mK@0°C and ~+100mK@100°C.
Overall the current version was designed to get the performance of the ADC, final version is targeted to get lowest cost/channel without sacrificing performance significantly.
Full project files will be released soon
TM