The ACIS focal plane (FP) temperature is a critical factor in the operation of the ACIS instrument on the Chandra spacecraft. The spectral response of the ACIS CCDs depends very strongly on this temperature. For this reason, the FP temperature is limited for each observation, depending on the spectral resolution required to achieve the science goals specified by the observer. To ensure to the fullest extent possible that the FP temperature is within the specified limits, the Chandra team runs predictive thermal models of the ACIS FP temperature for every observing schedule.
However, what if the ACIS FP thermistor was lost, as has happened for other thermistors on the spacecraft? In this case, the ability to accurately determine the calibration needed for a partiuclar observation would be compromised. For this reason, I have developed a machine learning model that predicts the ACIS FP temperature from other related thermistors. This is a simple LSTM model with a few layers, trained on the ACIS FP temperature and other thermistors from 2020-2025.
Below one can see an example prediction for the ACIS FP temperature using this model, against the actual temperature data (top-left panel), as well as the residuals vs. time (bottom-left panel). In the top-right panel, the residuals are compared to the temperature itself, the bottom-right panel shows the residual histogram. Currently, the model is able to predict the ACIS FP temperature with 1% and 99% quantiles of -0.8 $^\circ$C and 0.6 $^\circ$C, respectively.
