Future smart sensors will be able to utilize the maximum information content from data products, while minimizing the resources required to acquire, downlink, and process data. In-orbit calibration is required for space-borne radiometers in order to correct for gain fluctuations. Many sensors like radiometers are generally only able to produce calibrated scene measurements after reaching steady state. Waiting to reach thermal equilibrium to obtain useful data results in wasted power, excess useless data, and delays in obtaining useful data. Instrument power cycling provides a way to lower power use, but at the cost of pauses in data collection when the instrument is cycled off. Rapid power cycling can be used to reduce the average power draw of a radiometer, at the cost of increased measurement uncertainty. These power cycling techniques have been used on real systems, including the IceCube radiometer [1]. Using a convolutional neural network trained on synthetic data, a simulated radiometer can produce calibrated measurements with lower uncertainties and errors than conventional least-squares-regression (LSR) - based estimators. This approach presents an opportunity to reduce the average power draw of a radiometer by minimizing uncertainties of calibrated data products collected during rapid power cycling.