DOE Help: Hard-to-change factor drifts during experiment
May 27, 2019 7:11 PM(1334 views)
I am running a DOE where temperature is a hard-to-change factor. The scenario is a hot plate, on which the sample sits, and gets hit by a high-power plasma. The thermal conductance between the sample and hot plate is not constant, and is a function of the process chamber pressure. Normally, I deal with this by allowing a sufficient "pre-heat" time for the temperature to equilibriate regardless of pressure, before I start the process. The only temperature sensor is the hotplate control TC (i.e. can't measure actual temperature of sample).
Once I start the process (a high-power plasma), it tends to raise the temperature of the sample (and hot plate), causing a temperature drift during the run. However, I suspect that the drift is detected only when there is sufficient gas pressure (and therefore sufficient sample-to-hotplate thermal conductance) to allow the hot plate temperature sensor to pick up the temperature rise.
Any ideas on whether or how I can deal with this complicated limitation in my DOE? Thanks in advance!
Use the factor settings as usual but measure the temperature during the run and update the temperature value in the data table before you perform your regression analysis. The Whole Plot variable will track the random effect of re-setting the temperature via the hot plate and the measured temperature level will track the fixed effect of temperature changes.
Thanks for the reply.. Any way to account for the confounding effect of the pressure factor on the measured temperature (i.e. perhaps the temperature rise is captured only when there is sufficient thermal conductance between the hot plate TC and the sample)?
I think I understand your question. Let me restate it: the effect of Temperature might depend on the level of another factor, Pressure. If that is your hypothesis, then add a cross term to the linear predictor in your regression model, Temperature * Pressure. Such an effect on the response is known as an interaction. Each term of the model is tested in addition to the whole model.