[sdiy] Temperature compensation results

[Ian Fritz]

Hi Rene!

At 05:55 PM 6/8/2003, René Schmitz wrote:

>Hi Ian,
>
>>To start with, here's a trick that I haven't seen mentioned in any of the 
>>synth discusssions.
>>If there is a voltage offset in an expo converter, it will produce a 
>>temperature drift.  This is an "absolute" drift (all frequencies drift by 
>>the same ratio) as opposed to a "scale factor" drift (scales stretch or 
>>contract).
>
>I would say there is always an "offset". Since Vbe isn't zero.
>(Comparing to the use of the term offset in differential amps, where both 
>transistors are kept at the same base potential.)
>
>>The "trick" is to deliberately add an adjustable offset and use it to 
>>cancel other sources of absolute drift, e.g., from temperature 
>>dependences in the oscillator core.  The relation between offset and 
>>drift is that 1 mV of offset in Vbe gives -129 ppm/K drift.
>
>I fail to see how that would be any different than adding an external CV 
>to the CV summing node. (Or detuning via the panel tune pots.)
>(I would assume that you would need a voltage that is PTAT for that to work. )

That's the point!  The added offset I'm describing is *not* PTAT, whereas 
an added CV would be, because it would be multiplied by the tempco resistance.

The voltage you normally apply to the base (deliberately) is proportional 
to the input control voltage:  Vb = const  Vin.  If the constant is PTAT, 
then the converter is properly compensated.  If the constant is not PTAT, 
say just some fixed number, then the frequency response goes as exp{A Vin / 
T} where A is independent of T.  This gives a temperature drift, but it is 
a "scale factor" drift, because of the Vin factor.  In other words, the 
drift rate (d/dT)Ln(f) is proportional to Vin.

What I described in my trick is different. It is an offset voltage 
independent of Vin and of T, i.e., Vb = const.  In this case, f goes as 
exp{B / T}.  This gives a drift that is an "absolute drift", by which I 
mean (d/dT)Ln(f) is constant (independent of Vin).  This is then made to 
cancel other sources of absolute drift, such as the integrating cap's tempco.

>Could you please elaborate on the background of this?

I hope the above helps.  I've spent some time over the last month analyzing 
drift processes, including detailed analyses of ideal and non-ideal tempco 
resistors and of active compensation schemes.  Jim Patchell was the one who 
pointed out to me the importance of distinguishing between absolute and 
scale-factor drift.  I hope I can find some time to write all this up, as 
it is easy to get confused on.

>>I added this to my dial-a-tempco circuit, so that there are now two 
>>adjustments: one for scale-factor drift and one for absolute drift.
>>After proper adjustment, the total drift is now down into the regime 
>>where it is quite difficult to measure, but it is definitely below 50 ppm/K.
>
>Awesome results!

I agree!  I wasn't ever sure I would be able to get there, but it seems 
OK.  And of all the people selling VCO's who say they don't drift, where is 
there one willing to give actual numbers???

>>There has been an objection that this method would be difficult to use 
>>because of the many temperature cycles it would take to dial in the 
>>corrections.  Actually it is quite easy, if you do it correctly.  The 
>>correct method is as follows:  First, set the compensation voltages to 
>>zero and carefully measure the scale-factor and absolute drifts.  Then 
>>use calculated equations to figure out what levels of compensation 
>>voltages (or pot turns) are needed for correct compensation.  Finally, 
>>just dial in these corrections using a DVM (or by counting turns).  This 
>>should bring you very close to exact compensation, but it may be 
>>necessary to iterate once to get right on.
>>To measure the temperature drift I used a very simple "oven".  The 
>>circuit is mounted inside an aluminum chassis box with standoffs.  The 
>>box is heated with a drug-store heating pad wrapped around it.
>>Temperature is measured with an IC sensor mounted right next to the 
>>converter.  This is a simple, quick, quiet and reproducible method.
>
>Do you think that individual oscillators would have to undergo this 
>procedure, or do you think once you know the values you can skip the 
>measurement and just use those obtained on another specimen?

Boy, good question.  At this level I would hate to rely on batch 
consistancy for all the parts.  But probably you could get a good start by 
initially using the same settings and then tweaking a bit.

   Ian