Hi,
My thoughts.
Please mind: Low impedance signals are good, because they are stable, reliable and they are not prone to pick up errors like noise or leakage currents.
A thermocouple signal is a low impedance signal. A good signal.
But the first you do is to connect a series resistor of 10k.... this makes the signal weak....and prone to errors.
The same is with the Opamp output.
The first I'd do is to make both parhs more low impedance. Maybe 100R.
***
For sure this influences the filters you built.
Let's analyze them:
R23 / C18 form a low pass with tau = 100ms
R2 / C20 --> tau = 10 ms
R25 / C26 --> tau = 1000ms
Three different frequencies. Is there a reason?
I personally don't like electrolytic capacitors as filter capacitors. They are not suitable because they suffer from leakage currents and are not suitable for higher frequencies.
Now you may say there are no high frequencies. Only partly correct, because the ADC within the microcontroller causes current peaks at every conversion. These short peaks call for a fast capacitor.
My recommendation:
Use 100R / 100nF at the input and 100R / 100nF from Opamp to ADC. Both filters are just for noise suppression.
Then use the 1000ms filter at the Opamp feedback. But no electrolytics capacitor. This is the true signal LPF.
Best filter capacitors are foil capacitors, but here - because you don't process audio data - you could use ceramics capacitors.
They suffer from poor linearity, but with your DC style signal this should cause no problem.
Microcontroller:
Be sure to switch OFF the port pullup feature at the ADC inputs.
Switch the port as ADC input and leave it in this state.
Some additional hints:
My personal taste: I'd rather use a 100ms analog filter and add the 100ms filter as a software filter.
ARef: I'd replace the (slow) electolytics with a 1uF ceramics capacitor. I assume it results in better VRef stability.
AVcc: I don't like undamped LC filters. They cause a high Q resonance. I'd rather replace the L with an 100R and the C with a 1uF ceramics.
Klaus