AD588 self-heated reference may be possible

[magic]

I came up with this idea a few months ago after reading about the self-heated LM723.

The trick is very similar. AD588 has a number of on-die opamps. If we reverse-bias one input transistor of a typical opamp, the input stage current source will likely saturate and pin the input pair's emitters to the rail. Then, application of Vce(sat)+Vbe(on) to the other transistor's base will turn it on and control the VAS and the opamp's output. This voltage has a decent thermal coefficient and can be used to sense or stabilize die temperature.

Thanks to Noopy's die photographs, we don't need to guess which transistor to cut off and what's the polarity of the input stage. It's a µA741 input stage, so the VAS is controlled by the inverting input and the noninverting input is disabled by tying to the negative rail. This is the luckiest outcome we could hope for, because to the inverting input we can apply negative feedback and make the opamp buffer its own "input threshold voltage" on the output. Also important, amplifiers A3 and A4 appear to be completely independent from everything else and to contain their own bias generation circuitry, based on individual buried zeners and some ratio'd transistors. Playing weird tricks on either of those opamps shouldn't affect the other opamps nor the reference.

Due to the unusual input circuitry of the 741, its threshold voltage is actually higher. To turn the VAS on, we need to overcome the Vbe of the input NPN (Q2), the Vce(sat) of the input PNP (Q4), the combined Vbe of the Darlington VAS (Q16~17) and whatever voltage exists across its degeneration resistor (R11), determined by VAS standing current. All things considered, something in the vicinity of 1.8V and -6mV/°C is reasonably expected.

Yesterday I built a small board which wires amplifier A4 exactly that way and leaves all other pins floating. Initial results are optimistic: output voltage is indeed 1.7~1.8V at room temperature and I observed -8mV/°C drift over a 20°C excursion (I put it out the window for a few minutes). Since then, the device is running indoors and I plot its output against ambient temperature every now and then and it starts to look like a -8mV/°C slope again. There is still some 0.25°C p-p noise and hysteresis. I think I will glue a thermocouple to the chip's body, wrap everything in foam and try again. Hopefully this will get rid of it.

What will be the long term stability? No idea, but it's easy enough to test.
 

[magic]

Single supply conversion is another thing I wanted to try. If Vlow = Vs-, the chip's own output can be divided down to easily obtain a reference voltage for thermal stabilization. Furthermore, power supply complexity is reduced and there is no risk of damage if one rail goes out but parts of the circuit are still connected to ground. Unfortunately, simply shorting Vlow to Vs- causes it to source over 40mA of current. This is resolved by forcing A2 into negative saturation, but then A1 stops working and outputs some nonsense like 6.4V. I will try to understand what happens and see if a workaround is possible, but I'm not optimistic. It looks like A1 requires negative supply headroom below ground or shares some circuit parts with A2.

The backup plan is to generate a precision virtual ground. I breadboarded a circuit which achieves it using A2: Vhigh is divided 100k/50k and the resulting voltage goes to A2. AD588 powers up correctly in this configuration, Vlow becomes 5V and Vhigh is 15V.

I experimented a bit with self-heating. Simply applying 1.6V to A4 IN- with IN+ and OUT shorted to ground (single supply), by analogy to the LM723 circuit, doesn't work for some reason. I monitored die temperature using A3 and it went below 1.6V but A4 was still heating, until I turned the power off at 1.5V.

The next attempt was to setup A3 as a thermometer and route its output to A4, which compared it against 1.6V. This sorta worked, but there was high offset and poor regulation - simply blowing at the chip caused the temperature to drift a few degrees. I suspect it's because A4 was working at the very edge of its CMIR. Maybe it would help to amplify A3's output before feeding it to A4, but that's getting too messy for a breadboard prototype.

Time to make a board, I think. And complete reverse-engineering of the internal circuits, because there is too much guesswork in it at this point.

[Kleinstein]

The OP as temperature sensor works at the very edge of the common mode range - it kind of uses this edge as the temperature dependent signal.  To bring that signal to A4 one would need to add some level shift (e.g. mix in a little (like 0.5 V) from the reference).
The A4 part could than act both as heater and regulator part.
It is a funny idea to use the amplifier as temperature sensor and for regulation.

The AD588 is however quite expensive - at east as a new part.

[magic]

Well, I got mine from China

New ones aren't exactly cheap, but it's a nice IC which packs a lot of functionality in a small, plug and play package.

Meanwhile the mystery of most of my problems has been resolved. Turns out, A2~A4 use diamond buffer output stages and when the output gets within about two diode drops of either rail, the corresponding driver forces its active load into saturation. This in turn disturbs other bias current sources: saturation to the positive rail debiases the VAS, saturation to the negative rail debiases the whole amplifier.

Furthermore, A1 doesn't have its own bias generator, it's biased by a current mirror originating from A2. This means that bringing A2 output to the negative rail basically shuts down A1.

Solution for A2: never allow A2 output near ground, unless you don't care about A1. Use real or virtual ground, remember that reasonable common mode input range for A2 begins at 2V above the rail if not higher.

Solution for heating: connect the output to the rail through a few diodes or some TL431 or zener. I added TL431 to my breadboard contraption and now it works much better, and it works in the simplest configuration possible: fixed voltage applied to IN- of A4, and A3 is not involved in regulation at all. Of course I configured A3 for temperature monitoring to verify performance. Now I can blow at the chip or even touch it and it stays pretty constant, only current consumption increases

[David Hess]

It will not work for your application but if you have access to the offset null pins, then the operational amplifier can be configured to amplify its own input offset voltage to make a pretty good temperature sensor.

[magic]

Not sure if I follow. Of course input offset voltage can be amplified, even without the nulling pins, but offset and offset drift are small, unpredictable and possibly zero. Maybe you meant something else?

CMIR, or output saturation voltage for EF outputs, are theoretically expected to exhibit drift and the coefficient should be somewhat repeatable from sample to sample.

By the way, with more accurate thermal coupling, the actual drift of my chip is turning out to be almost exactly -7mV/°C. It's not very important, but I want a way of estimating die temperature to avoid killing it.

[magic]

By the way, I now have a full schematic of AD588 rev. A and bits of rev. C. Maybe somebody will find it useful or at least amusing


The reference circuit is the same concept as AD587. Revision A has resistor R9 to trigger startup, revision C replaces it with the startup circuit known from AD587.

A1 is a very simple amplifier, redesigned in rev. C. Both versions require internal bias voltage from A2.

A2~A4 are modified 741 amplifiers with diamond buffer outputs and high-tech biasing. A2 can only sink on rev. A and C, but my chip also sources, so perhaps there is yet another revision.

The problematic transistors are Q16 and Q17. When the output swings negative, Q17 emitter can almost reach ground, reducing current and voltage at R7. Then R7 and Q16 pull down on Q23, which drives the bias rail.

Rev. C adds bias current cancellation on GND SENSE+ pin, active current limiting on PNP outputs and BE breakdown protection for Q17. Rev. A appears to lack the protection.

[David Hess]

Not sure if I follow. Of course input offset voltage can be amplified, even without the nulling pins, but offset and offset drift are small, unpredictable and possibly zero. Maybe you meant something else?

Deliberately creating an input offset voltage by using the offset null pins also creates a predictable input offset voltage drift.  Precision operational amplifiers take advantage of this to null their input offset voltage drift when their input offset voltage is nulled.

[magic]

Seems reasonable. That being said, you could simply monitor the average voltage at the null pins and it's likely to be PTAT or otherwise drifty. You get to use the opamp too, as long as it never saturates.

[magic]

 

Solution for heating: connect the output to the rail through a few diodes or some TL431 or zener. I added TL431 to my breadboard contraption and now it works much better, and it works in the simplest configuration possible: fixed voltage applied to IN- of A4, and A3 is not involved in regulation at all.

So that works up to a limit. When the heater turns off for whatever reason (and it does after a few minutes of undisturbed operation) it goes to the negative rail and stays there forever. Unlatching it is as easy as briefly grounding IN- or raising IN+, so I wonder if there is a simple way of doing the latter automatically when the output drops below 2V...

Alternatively, I could use two opamps: negative feedback thermometer with 2x gain and comparator/heater. With TL431 in the heater's output, the loop clearly has too much gain and turns into PWM at a few Hz. How does one compensate an oven?

A lazy way is to remove the TL431 and hence debias the heater. Temperature reading seems more stable, but I haven't scoped the supply current yet so not entirely sure. Seems to pass the finger test with only a brief disturbance. There is tens of mV input offset voltage in that mode, maybe not a problem if it remains stable. But I'm getting tired today...

[RandallMcRee]

TL431 oven controller (scroll down):

http://techlib.com/electronics/ovenckts.htm

Used this several times. Recently did a temperature measurement of a hammond box with one of these inside and got 35 degrees C, std dev 0.01 degrees--after one hour of stabilization.

[magic]

I'm starting to have second thoughts about the "self" aspect of self-heating, when it comes to chips that weren't designed for it.

Everybody has heard of thermal gradients across integrated circuits, but just how bad can they really get in practice? Well, with AD588 and its three fully independent opamps on one die we can try to find out

The picture below shows location of the small signal (temperature sensing) and output (heat generating) stages of A2, A3, A4.


I used the following test conditions:
- single supply, 15V to be safe with short circuits to ground
- all amplifiers wired as buffered thermometers: unity gain and IN+ open circuit (too lazy too ground them for a quick test)
- outputs optionally shorted to ground (50mA current, 13V dropped inside the IC → 650mW) or loaded with 100Ω (18mA, 13V → 230mW)

In quiescent state, differences between the outputs of different amplifier are a few mV and quite stable.

A4 dissipating 230mW: A2, A3 difference increases 3.5mV or 0.5°C
A4 dissipating 650mW: A2, A3 difference increases 11mV or 1.5°C
A2 dissipating 650mW: A3, A4 difference increases 2mV or 0.3°C

After removal of load from the output, the difference returns to normal very fast, "feels" like the time constant is less than 1s. I think it's a sign that this is indeed a result of thermal gradient, rather than something innocent like different thermal coefficient in each amplifier.

Notably, A2 input stage is located just next to the buried zener, which means that the offset above represent a realistic drift of A3 temperature wrt zener temperature when the far end of the die varies its dissipation.

At last, we can take a quick glance at how true self-heated zener ICs deal with thermal gradients.
LTZ1000 - symmetric reference structure surrounded by a circular heater
LM399 - heater along one edge, all temperature critical bits including heater sensor arranged in one equidistant row

There we go. There was probably a reason for that. Damnit.

[Kleinstein]

The thermal conductivity of silicon os quite higher. So the temperature gradients are usually not that large.
Ideally one would not like much heat directly on the chip, but more like only a small fraction (like 10%) for fast reaction and most of the power (e.g. especially the I part of a PID regulator part) externally to compensate for a variable external temperature.

The important part is to have the sensor on chip, not so much the heater.

[Gyro]

Ceramic Dip packages are very amenable to having smd heater resistors glued to the outside, put in a thermally insulated container and the pins air-wired.

If the die geometry isn't designed for balanced heating, warming the package should give much better thermal gradients (and better control of the package thermal mass, which an on-die heat source would be having to drive anyway).