Hello, I am a little bit confused when analyzing this design.
In fact, after following the path of the sensor output signal (TEMP PWM), I see that it was conditioned (impedance matching by U20B;amplification by U20A) than transmitted, after another amplification, to the feedback pin of the IC. But the signal temp in transmitted to the micro-controller (in another sheet) is simply the "image" of the sensor output, not a signal obtained from a real current measurement. Is that technique widely used in temperature control with Peltier elements?
View attachment 333019

 

No, Peltier effect is very unsatisfactory used as T sensor, because all Peltiers are produced for work with abnormally high temperature flow DENSITY on area, thus the impact of cold side (or contrary warm side) will be overwhelming to destroy every least accuracy. For moderate temperatures are made the diode type ICs with analog or even digital output in about 0.05-0.1 C metrologically warranted accuracy and about 10 microCelsius the sensitivity. For higher temperatures as Silicon may stay alive, there are plethora of thermocouples and for temperatures over some 1500-3000 the platinum have a best place to be used. If any more - before Belorussia become the tyran reigned inquisition country, they had a wonderful construction of Mass Flow Institute able to measure with Planck law about 0.1 C accuracy. The paradigm was to measure not the object in the oven but the standard clay piece thus always giving the colour being precisely known. And other quirk was they used not the total infrared flux but two filtered wavelengths, one beyond the peak and other over the peak, thus being able to calculate precise peak maximum wavelength. Thus they used actually not a Planck law but the Wien second law instead. Now, seems all that work is completely destroyed and I am suspicious he author may happen be prosecuted as he was friendly to cooperation and not much secretive.
Meanwhile in the most simple cases when +/- 1 C is enough, the Arduino may measure itself, having a genuine sensor onboard. Yet here is one surprise to me. I used for long time the Atmega 328 in simple devices because it cost only 0.99 USD already mounted, soldered in full complectation, whilst the Rasp costed 15-30 USD. Nowadays it all had been shifted feet up, head down. Specific most minimalist Raspberry model with two cores cost even 0.75 USD, while two cores changes many things cardinally.