Measuring mK with small RTDs

[Anders Petersson]

PT100 and PT1000 are stable over time but the small signal level is a challenge for ADCs, if mK resolution is needed. Especially for small sensors, as they have high self-heating.Consider a nominal PT100:
1 K = 0.385 ohm change in resistance
1 mK = 0.385 mOhm

At the common excitation current 1 mA, small sensors face a large heating effect that must be dissipated with good thermal coupling. Air is not good thermal coupling. For example, Heraeus 32208550 (https://www.mouser.se/datasheet/2/619/m_222_e-1919962.pdf) heats up 0.4 K/mW and the datasheet helpfully reminds of the self-heating effect. At 0 C, the effect is 1 mA through 100 ohm = 0.1 mW, so the self-heating is 0.4 K.

0.1 mA would decrease the self-heating by x100 to an acceptable 4 mK. But at 0.1 mA, a 1 mK change gives only 0.385 mOhm * 0.1 mA = 0.0385 uV signal change. Even a cursory analysis says this is far below the noise floor, with thermo-EMF ca 1 uV and regulator noise 1-10 uV, just to name a few noise sources.

Even a 1 mA excitation current gives only 0.385 uV/mK, probably requiring quite some averaging time to extract 1 mK resolution. That's OK for large sensors, but ruins the idea of a short pulse of the excitation current to reduce the 0.4 K self-heating.

I'm drawing the conclusion that with PT100, we can't both have high accuracy (high resolution and low self-heating) and fast response time (small sensor with weak thermal coupling).

TI app note https://www.ti.com/lit/an/sbaa275/sbaa275.pdf doesn't give any help.

I've heard that measuring with both positive and negative bias currents can cancel out thermo-EMF, and there's also offset compensation by measuring the voltage without bias current. I haven't seen any circuit for these.

There's a commercial radiosonde "Vaisala RS41" that moved from the commonly used thermistor to PT1000, and they use a ring oscillator. Here it is reverse engineered: https://github.com/bazjo/RS41_HardwareI guess that can give high resolution, and maybe also cancel out thermo-EMF? They do have two reference resistors, which should be a clever way for a two-point correction. Radiosondes must be cheap, have fast response time and they experience a huge shift in ambient temperature so that company is probably on to something. Unfortunately I don't know enough analog electronics to recreate their solution.

Does anyone else know a good RTD conversion circuit?

[branadic]

Compared to SPRTs platinum or wire wound platinum temperature sensors (IPRT), thinfilm thermometers can show quite some large hysteresis, due to the cte mismatch of the platinum layer and ceramic substrate that results into strain.

Also uncertainties quite differ between different fabrication technologies*:

SPRT: u ≤ 10 mK
wire wound platinum sensor (IPRT): 10 mK < u ≤ 30 mK
thinfilm sensor (selected and preaged): 20 mK < u ≤ 50 mK --> not used as primary or secondary standards
thinfilm and wire wound sensors in general: 50 mK < u ≤ 200 mK

(*source: imetrologie, "Hysterese-Erscheinungen bei Platin-Widerstandsthermometern"; Metrologietage Kassel 10/2020)

-branadic-

[KT88]

A few thoughts on that topic:
- Self heating depends also on the thermal resistance to ambient (the media measured). The selfheating specified in the DS is against air. Having a good thermal coupling would significantly reduce the selfheating efect.
 - The stability of the electronics depend a lot on the effort that is allowed. If you aim for highest precision at a (almost-) constant temperature DC stability better than +/-100nV is doable.
- If the allowed effort is limited a Pt1000 has lower heat dissipation (I2R) and higher output voltages. Ibias influece and resitor noise are more noticable though...
- Another method that could be implemented at reasonable effort would be AC excitation. The narrow band signal togethwe with thermal EMF cancelling could give good results. Some ADCs support this approach like the AD7195.

Cheers

Andreas

[Dr. Frank]

Hello Anders,
at first you should make yourself aware, what requirements you really have in your temperature measurement problem:

- do you want to measure the absolute temperature very precisely? E.g. measure 300K with +/- 10mK uncertainty
- or do you want to determine temperature changes with high resolution? e.g. measure 1mK change @ 300k, absolute uncertainty (vulgo: 'accuracy') not important
- at which absolute temperature do you want to measure? e.g. at 100K, 200K, 300K, 400K, etc.
- how is the setup of your thermodynamic system? e.g.: measure in air, liquid, solid like a metal block, sink or source heat

Only then, specific answers to your questions can be given.

In short, a PT100 can be used for temperature changes to about 1mK, and absolute uncertainties (with calibrated sensors) to about 10mK. Expecting smaller numbers is useless.

You only need a device which contains a precise 1mA or 100µA current source, which can be switched on/off or reversed, and a stable high resolution voltmeter with 100nV resolution, or better.
Commercial DVMs with OFFSET COMPENSATION will do the job, like 34465A, 34470A and 3458A from Keysight, or 7510 and alike from Keithley, or Flukes 8508, 8588 and alike.

https://www.lakeshore.com/ offers calibrated sensors and specialized test equipment, I propose that you read there about temperature measurement techniques, that might clarify your requirements.

DiY solutions are possible, but tough to realize, as you need to mimic the resistance mode features of the DVMs mentioned above.

Frank

[KT88]

Another alternative to resitance measurements with DVMs or similar would be a bridge configuratin. This would allow for ratiometric measurements with AC excitation. This type of measurement would only rely on stable resistors. Some PWW resitors might be challenging to use with AC excitatin though. Other resistors like Vishay precision resistors and dividers would be fine for that purpose. The beauty of ratiometric measurements  is that it doesn't rely on the accuracy of a voltage reference.
An example is shown in the DS of the AD7195 on p. 43.

Cheers

Andreas

[MiDi]

You got the point why RTDs are very difficult to handle with in mK region - especially the Pt100.
1mK corresponds to ~4ppm change in resistance, which would require expensive resistors for a current source or 4W bridge alone to get low enough TC and drift for maintaining 1mK accuracy.

E.g. a Fluke 8588A would only be good enough to maintain <10mK accuracy with 100µA in 1kOhm range:
24h: 10ppm + 2mOhm
1yr: 16ppm + 2mOhm


Ingredients for a circuit could be:
-low noise & drift 100µA current source
-ADC with specs similar to those in 3458A/8588A
-ultra low noise, ultra low T-EMF precision frontend with gain ~1000x (depending on ADC input range and max. temperature) with specs similar to 34420A frontend or EM Electronics A22.

If you have no absolute need to, this is nothing you want to design

Depending on your needs a precision thermistor or digital sensor could be a much cheaper alternative e.g. the Ti TMP117 has drift of ~0.03K/300hr@150°C and noise ~15mKpp (8s averaging).

[KT88]

In a bridge configuration one half bridge could be an integrated divider which mainly relies on a stable ratio (TCR). Only one resistor must be a very high precsion (cost) one.
The eval-board of the AD7195 is $59. Resistors might be in the ballpark of additional 100 bucks -that's not to expensive if you really want to go to town with temperature measurement...

Cheers

Andreas

[ch_scr]

For historical context, here are a few quick shots of "AC Cryo Bridge R441" from CSAV "Československá akademie věd" - "Czech Academy of Sciences". I got device cheap and defective, please excuse quick hack to replace defective UDSSR opamp with uncommon pinout. Device uses CD4066 as analog switch for synchron rectification.

[arcnet]

You got the point why RTDs are very difficult to handle with in mK region - especially the Pt100.
1mK corresponds to ~4ppm change in resistance, which would require expensive resistors for a current source or 4W bridge alone to get low enough TC and drift for maintaining 1mK accuracy.

E.g. a Fluke 8588A would only be good enough to maintain <10mK accuracy with 100µA in 1kOhm range:
24h: 10ppm + 2mOhm
1yr: 16ppm + 2mOhm


Ingredients for a circuit could be:
-low noise & drift 100µA current source

Yes, quite low noise but nothing really low noise. Low drift is not required as long as the current source is stable enough during a measurement

-ADC with specs similar to those in 3458A/8588A

Yes, mainly linearity but this depends on the temperature range that needs to be measured. Some off the shelf ADCs are linear enough if the range is low enough/the (amplified) signal is in the right range. If this is not sufficient there are a couple of methods for measuring/linearizing ADCs e.g. "High-Performance ADC Linearity Test Using Low-Precision Signals in Non-Stationary Environments", Jin et al., 2005

-ultra low noise, ultra low T-EMF precision frontend with gain ~1000x (depending on ADC input range and max. temperature) with specs similar to 34420A frontend or EM Electronics A22.

Not really needed. A PT100 with 100 uA excitation current "generates" around 38.5 nV/mK which can be measured (usually fast enough) with off the shelf ADCs like AD7177, ADS1262/1263, ADS1281-1284 or MAX11216...
1/f noise and T-EMF: Current reversal with or without substitution technique as used by Isotech, delta method... But we can always increase the requirements e.g. high temperature RTDs with a base resistance of 0.25 Ohm and ~9 nV/mK @ 10 mA

Depending on your needs a precision thermistor or digital sensor could be a much cheaper alternative e.g. the Ti TMP117 has drift of ~0.03K/300hr@150°C and noise ~15mKpp (8s averaging).

Or some ultra stable NTCs as tested in https://ieeexplore.ieee.org/document/7281460 "The results show that the SMD type sensor from Murata manufacturing (NCP15XH103D03RC), intriguingly, is the most stable sensor among the sensors tested with a drift rate of 0.492 mK/year peak-to-peak." (test was done at room temperature 22 °C)

If you have no absolute need to, this is nothing you want to design

And definitely not cheaper to design than buying a thermometry bridge incl. SPRTs, fix-point cells, reference resistors and all the maintanence equipment

But the original question was about resolution not necessarily accuracy...
"I'm drawing the conclusion that with PT100, we can't both have high accuracy (high resolution and low self-heating) and fast response time (small sensor with weak thermal coupling)."

If fast response time/small elements are needed see https://netsushin.co.jp/en/product2.html or https://netsushin.co.jp/en/product1.html
The smallest ceramic wire wound PT100 element has a diameter of 0.4 mm (probe 0.5 mm), length 1 mm. Depending on the requirements they have even smaller 0.4 mm x 0.5 mm RTDs but with only 10 Ohm base resistance https://netsushin.co.jp/en/product8.html (4-wire configuration and still a wire wound)

Self heating can be eliminated by using pulsed excitation and/or measuring the RTD with two different currents and calculating the zero-current resistance.
The optimization of self-heating corrections in resistance thermometry, Pearce, 2013
https://www.researchgate.net/publication/258263469_The_optimization_of_self-heating_corrections_in_resistance_thermometry

Additional notes:
Keitley Low Level Measurements Handbook
https://download.tek.com/document/LowLevelHandbook_7Ed.pdf
and the very good
High Accuracy Electronics A collection of monographs on inductive voltage dividers, ratio transformers and related circuits plus some nice maths by Christopher I. Daykin that were posted here in the forum
https://drive.google.com/drive/folders/1TbomXeoBbIe-IVADaOOmyyLH9le-3wAt

[thermistor-guy]

PT100 and PT1000 are stable over time but the small signal level is a challenge for ADCs, if mK resolution is needed. Especially for small sensors, as they have high self-heating
...
There's a commercial radiosonde "Vaisala RS41" that moved from the commonly used thermistor to PT1000, and they use a ring oscillator.
...

Interesting. Recently, after crunching the numbers for an application - self-heating, sensitivity - I went the other way, from a PT1000 to a 100K glass-encapsulated thermistor.

I had also considered using a temperature-sensitive quartz crystal, like an Epson HTS-206, but decided against it for now. It would be a project in itself, to screen and characterize a batch of devices for good aging, hysteresis, stability, etc. I have those crystals on hand, though, and want to return to this one day. So many projects, so little time ...

[maat]

As other people pointed out before, it can be done and you don't need any special voodoo parts or custom ADCs. I did do similar design before. Before I go into details, let me ask, why you insist on using a PT100. Do you need absolute uncertainty?

If you spend some decent money on a 10k thermistor you can get away with less hassle. The good ol' YSI thermistors are great (back then sold to Measurement Specialties and now TE). The 44000 series claims better than 10 mK over 10 months (http://www.farnell.com/datasheets/169207.pdf), it is repoerted to be even better. J. Dratler Jr. claims <= 100 µK in 8 months for a 4400 (https://aip.scitation.org/doi/10.1063/1.1686523). They can be had for around 10 € (https://www.mouser.de/ProductDetail/Measurement-Specialties/701036?qs=tiyUwePmMSx745wk1vI7CQ%3D%3D).

Now, for the resistance 'bridge'. I am calling it 'bridge', because it is not a classic resistance bridge, but rather transferring the resistance via an ADC. I used an LTC25008-32 with two LTC2057 input buffers (The OPA189 doesn't like large input impedance, unfortunately). The voltage reference is an LTC6655-4.096 with some moderate additional filtering. The reference is fed into both the ADC and the current sink (yes, sink!), so don't worry about the tempco, it cancels out. The current sink is again filtered and bootstraped using a JFET and finally guarded using an OPA827 (this makes it a positive current source again). The current source polarity can be switched using a MAX329 mux. The current source and ADC including the front end are under a shielded cap, because air draft will cause a headache. Finally there is a relay on board to switch the frontend to a reference resistor for autocalibration (VHP101).

The ADC is runnig at 1 MHz clock and the current source polarity is cycled in between measurements. Linearity is < 1ppm after calibation. Tempco < 0.1 ppm/K. Your only problem is humidity, so you need hermetic resistors and clean PCB or frequent ACAL. THe cables are another soruce of error, if you are cycling the current. If the cable show dielectric absorption, that will be pop up as additional noise. I am using a PTFE dielectric.

The frontend is designed for 10k thermistors, which I typically run at 50 µA. I do not need any amplification. If you need amplification you must most likely bootstrap the amplifiers, but requirements for the resistor network is fairly moderate. Keep the tracking TCR below 1 ppm/K and you should be fine. This setup will put you at the physical resolution limit as calculated by Larsen (https://aip.scitation.org/doi/10.1063/1.1683078), which is about 3 µK RMS, but only if you use a fairly large thermistor, none of that tiny rubbish. All of that will put you in the ballpark of about 350 € / channel.

Finally, I have attached a 12h measurement of that setup. It is actually a two channel unit. One channel is measuring the internal reference (CH2), the other is connected to a 10k reference resistor on the bench, hence the higher noise. I have attached the same sample twice. Once in Volts and the other in Ohms. The sampling rate is about 20 / min, the filter used is a butterworth filter with a cutoff at 1/10 fc.

RMS noise is about 1.5 mΩ (CH1) or 0.77 mΩ (CH2), which be equivalent to 3.8 µK or 1.9 µK.

I can take a few photos of the board tomorrow, when I am back at uni.

[David Hess]

I've heard that measuring with both positive and negative bias currents can cancel out thermo-EMF, and there's also offset compensation by measuring the voltage without bias current. I haven't seen any circuit for these.

The idea is to apply AC excitation, usually by chopping, and then synchronously demodulate the output.  Essentially the sensor is included within the chopping loop that makes a chopper amplifier so accurate.

This results in removing the DC offset caused by thermocouple effects, and synchronous demodulation removes noise by using a very narrow bandwidth.  And since the measurement is made at AC instead of DC, flicker noise is removed as well.

0.1 mA would decrease the self-heating by x100 to an acceptable 4 mK. But at 0.1 mA, a 1 mK change gives only 0.385 mOhm * 0.1 mA = 0.0385 uV signal change. Even a cursory analysis says this is far below the noise floor, with thermo-EMF ca 1 uV and regulator noise 1-10 uV, just to name a few noise sources.

A 0.0385 microvolt signal change is 38.5 nanovolts which would be difficult at best to measure with a DC bandlimited amplifier because of flicker noise, but a moderately low noise amplifier could achieve 4 nanovolts rms in a 1 Hz bandwidth at a slightly higher frequency, considerably easing this measurement.

[Anders Petersson]

Thanks for all replies so far! I barely have time to research one reply before the next comes.

KT88 and arcnet mention AC excitation, example AD7195 datasheet p. 43 (attached for easy access). That looks plausible. Is that called an AC resistance bridge? But I think whetstone bridges are incompatible with 4-wire measurements, so the sensor must be very close to avoid errors from cable resistance.
Arcnets suggestion with two excitation currents is also a good idea.

at first you should make yourself aware, what requirements you really have in your temperature measurement problem:

I'm open to compromises, but ideally I want to measure in -50 C - +70 C air.
I guess a solution might need to prioritize either low absolute uncertainty at long integration time, or prioritize relative changes at high resolution and high sampling rate. I would be interested in either, although thermistors are looking good for the latter option.

DiY solutions are possible, but tough to realize, as you need to mimic the resistance mode features of the DVMs mentioned above.

Yes, this challenge is for a DIY solution.

Compared to SPRTs platinum or wire wound platinum temperature sensors (IPRT), thinfilm thermometers can show quite some large hysteresis, due to the cte mismatch of the platinum layer and ceramic substrate that results into strain.

Also uncertainties quite differ between different fabrication technologies*:
[...]

Good point that PRTs have their flaws too. (I'm not sure what "uncertainty" means in this context though.)

Yes some thermistors are as stable over time as thin-film PRTs... I think I've seen https://ieeexplore.ieee.org/document/7281460 before but it's behind a paywall now(?). Is it available somewhere? Arcnet quoted SMT part NCP15XH103D03RC. SMT would be challenging to use since I don't know how to insulate the sensor without risking the stability.

[maat]

I'm open to compromises, but ideally I want to measure in -50 C - +70 C air.

I don't want to rain on your parade, but once you get below 10 mK, you definitely need an enclosure, better to have multiple layers as well. And in that case you might as well measure the wall temperature of the innermost enclosure. You will see excursion of about +-10 mK in aluminum block if you walk past it at meter distance, because moving air is very much non-isotropic.

[thermistor-guy]

...
Yes some thermistors are as stable over time as thin-film PRTs... I think I've seen https://ieeexplore.ieee.org/document/7281460 before but it's behind a paywall now(?). Is it available somewhere? Arcnet quoted SMT part NCP15XH103D03RC. SMT would be challenging to use since I don't know how to insulate the sensor without risking the stability.

BIPM (2014) recommends glass encapsulation, if using thermistors for thermometry.

Rudtsch and Rohden, PTB (2015): "It has been demonstrated that with an NTC thermistor in a temperature range between 0 °C and 60 °C, similar uncertainties can be achieved as with Standard Platinum Resistance Thermometers."

[Echo88]

Your measurement results are impressive maat, please share more of your design.
The description of your guarded current source sounds much like the one used in the Isotech MicroK, with its Improved Howland Current Source, Cascode Stage and Guard Stage: https://isotech.co.uk/wp-content/uploads/2020/09/substitution-measurement-topology.pdf

@Anders Petersson: https://sci-hub.se/10.1109/ISIE.2015.7281460

[TiN]

Here is interesting design too : https://aip.scitation.org/doi/10.1063/5.0035673

[branadic]

I would like to point another approach using either PT1000 or 1k NTC such as Murata's low drift series, combined with a resistance to digital converter based on TDCs. Why?

1. It uses AC excitation, to charge up a reference capacitor such as 10 nF C0G/NPO or better to 3.3 V and discharge it alternating through the PT1000 and a high precision reference resistor, while measuring discharge time with about 22ps resolution, so it is already a ratiometric measurement too

2. Assuming a threshold voltage of 1.65 V at the internal comparator and 2 measuremens per second with 2 (alternative, none, 4 or discharge time measurements for averaging, the following calculated power losses inside the temperature sensor can be found:

--> (3.3 V - 1.65 V) *10 nF = 1.65 V * 10 nF = 16.5 nC
--> 16.5 nC * 2*1/s * 2 = 66 nA
--> 1000 ohm * (66 nA)^2 = 4.356 pW

Advantage: The results of such principle are already digital, no need for an analog frontend, constant current source or ADC, just a single chip with some passives around it.
Just my 2 Cents

-branadic-

[KT88]

KT88 and arcnet mention AC excitation, example AD7195 datasheet p. 43 (attached for easy access). That looks plausible. Is that called an AC resistance bridge? But I think whetstone bridges are incompatible with 4-wire measurements, so the sensor must be very close to avoid errors from cable resistance.

Good point. The equivalent to a 4-lead measurement is the 6-lead measurement where you have force- and sense leads for the bridge supply. The sense leads will be connected to the reference input of the ADC.
The advantage of this approach is the sole dependancy on the resistor ratio(s).
Some ADCs support a ratiometrich measurement of a half-bridge as well. This could be both by connecting the passive resistor to the reference input or by multiplexing between the RTD and the passive resistor and calculating the ratio in the digital domain. A simultanious sampling ADC is preferred in that case although not nescesary if sufficiant filtering in front of the ADC is possible.
Using the reference input has the advantage of having an ADC reading that is inherently proportional to the resistance ratio.
Edit: It is more a chopping technology. It averages out errors due to thermal EMF (the polarity of the thermal EMF is inverted in relation to the bridge voltage and output).

Cheers

Andreas

[tszaboo]

Step 1 is to forget about PT100, and select a PT1000. Which is not that obvious, as best off the self Class 1/10 Din still has 30mK accuracy in the best conditions. And then, ask yourself, why are you measuring air with 0.001K accuracy? At work we have a dozen PT1000 on the office desk, they are going to measure something like 0.3K difference without any sort of trick involved.
And then you need a reference resistor. For mK, in typical environment, you need what? Below 1 PPM temperature stability?
Self heating is a problem? Yes, then turn off the excitation current when not measuring. And calculate the thermal capacity of your probe. You can take 24 bit accurate samples in 0.1 sec, the settling time of your filters matter more than the ADC.

Yes, there are good RTD circuits, they are listed in the application note you linked.

[Dr. Frank]

at first you should make yourself aware, what requirements you really have in your temperature measurement problem:

I'm open to compromises, but ideally I want to measure in -50 C - +70 C air.
I guess a solution might need to prioritize either low absolute uncertainty at long integration time, or prioritize relative changes at high resolution and high sampling rate. I would be interested in either, although thermistors are looking good for the latter option.

Hello Anders,

obviously you pose a physics problem here, not an electronic measurement problem.
For your thermodynamic needs, measurement of the air temperature, a simple and cheap NTC thermometer with 0.1°C resolution and 0.5°C accuracy is fully sufficient.
Anything better is physical and engineering overkill.

If you would really need a precise measurement in a solid state system, for example the determination of a physical phase transition of a High T_c - superconductor, or on temperature fix point thermometers, you really would need such 1mK resolution or 10mK uncertainty.
This would also require a proper, solid thermodynamic setup with thermal anchors, calibrated sensors (diode, NTC, PT100, CGR, ...) and a profound investigation / calculation of heat flows and alike.

I still recommend to use one of the better bench DMMs (which includes PT100 temperature conversion) and a medium grade PT100, and you're done.

Frank

 

[Kleinstein]

I would like to point another approach using either PT1000 or 1k NTC such as Murata's low drift series, combined with a resistance to digital converter based on TDCs. Why?

1. It uses AC excitation, to charge up a reference capacitor such as 10 nF C0G/NPO or better to 3.3 V and discharge it alternating through the PT1000 and a high precision reference resistor, while measuring discharge time with about 22ps resolution, so it is already a ratiometric measurement too

2. Assuming a threshold voltage of 1.65 V at the internal comparator and 2 measuremens per second with 2 (alternative, none, 4 or discharge time measurements for averaging, the following calculated power losses inside the temperature sensor can be found:

--> (3.3 V - 1.65 V) *10 nF = 1.65 V * 10 nF = 16.5 nC
--> 16.5 nC * 2*1/s * 2 = 66 nA
--> 1000 ohm * (66 nA)^2 = 4.356 pW

Advantage: The results of such principle are already digital, no need for an analog frontend, constant current source or ADC, just a single chip with some passives around it.
Just my 2 Cents

-branadic-

The TDC method only works well if the capacitance is also very stable - so OK on the PCB, but a problem when using cables. Also the switch resistance can be a problem.

AC excitation is prefered, but not that fast.  A NTC can be nice for a small range, but limited if the temperature range is large.

[branadic]

Kleinstein, I expect your response is just speculation but not based on experience? Move the reference resistor to the temperature sensor, use shielded and twisted cable, this way influence of the cable already cancel out. As I also wrote, the cap should be of type C0G/NP0 or better, such as foil capacitors
The AC excitation and switch is part of the chip, so no worries here. To name it, you can use the PCap0x or PicoStrain family, was done and proven to work before and application notes on that are available too

-branadic-

[Anders Petersson]

It will take me hours to read all papers and ideas posted here so I'll just write a small note for now -- the question "why?". I need to do an initial adjustment of many fast-response thermistors in the production of radiosondes and sensors for drone-based atmospheric research. This leads to several in-house use cases. For example, to characterize the variation in temperature across the volume of my calibration chambers and the fluctuation over short time spans (>= 10 Hz). Of course, I also need reference thermometers that can go a year between traceable calibration. I'd be happy to discuss this fascinating area more but I don't want to change the topic of this thread.

I'm already coming to the conclusion that fast-response 1 mK measurements is best done with small thermistors. For the use case of characterizing the chambers, the sensors need to agree with each other more than they need absolute accuracy. The traceable reference thermometers don't need that fast response time and are best realized as bigger sensors that are insensitive to self-heating.

None of the above forces me to build a DIY RTD converter, so my question is actually mostly academic. If there's a schematics a non-expert like myself can use, I might CAD and build a few as a more portable alternative to bench DMMs, but it's not worth my time to research a design myself.
The easiest solution I spotted so far is:

Self heating can be eliminated by using pulsed excitation and/or measuring the RTD with two different currents and calculating the zero-current resistance.
The optimization of self-heating corrections in resistance thermometry, Pearce, 2013
https://www.researchgate.net/publication/258263469_The_optimization_of_self-heating_corrections_in_resistance_thermometry

Quoting the summary of that paper:
"It is shown that the optimum solution is to employ two measuring currents in the ratio 1 : 2 and allocate the measurement times at the lower and higher currents in the ratio 8 : 1."

However, then I have the problem of generating currents of exactly ratio 1:2, instead of the simple voltage regulator + ratiometric measurement I picked from the TI app note. Precision current sources don't look that easy to design. Ideas?

[KT88]

If I got it right, your goal is a very good repetability of the calibration reslts for your thermistors...
In that case there is probably no viable way around an oil bath. Yes its a messy pain in the neck.
Reasons:
1. In air you need to control humidity and flow to get fairly stable readings.
2. The thermal resitance against air is much higher, emphasising the effect of self heating.
3. Tiny differences in volume/surface area of the thermistors will have a noticable effect on heat transfer.
4. It will be pretty tough to control thermal gradients in air/gas ( it's not easy in liquids already).
5. Your reference sensor can have a fairly large heat spreader, reducing the temperature difference due to self heating significantly. A large sensor would help as well.
6. Residual heat gradients will show a very good repeatability.
7. Humidity can be kept out of consideration (mostly). The repeatability would likely benefit from pre-conditioning the thermistors upfront (gentle baking).
 
Cheers

Andreas