Search

The largest source of noise in exoplanet and brown dwarf photometric time series made with Spitzer/IRAC is the coupling between intra-pixel gain variations and spacecraft pointing fluctuations. Observers typically correct for this systematic in science data by deriving an instrumental noise model simultaneously with the astrophysical light curve and removing the noise model. Such techniques for self-calibrating Spitzer photometric datasets have been extremely successful, and in many cases enabled near-photon-limited precision on exoplanet transit and eclipse depths. Self-calibration, however, can suffer from certain limitations: (1) temporal astrophysical signals can become aliased as part of the instrument model; (2) for some techniques adequate model estimation often requires a high degree of intra-pixel positional redundancy (multiple samples with nearby centroids) over long time spans; (3) many techniques do not account for sporadic high frequency telescope vibrations that smear out the point spread function. We have begun to build independent general-purpose intra-pixel systematics removal algorithms using three machine learning techniques: K-Nearest Neighbors (with kernel regression), Random Decision Forests, and Artificial Neural Networks. These methods remove many of the limitations of self-calibration: (1) they operate on a dedicated calibration database of approximately one million measurements per IRAC waveband (3.6 and 4.5 microns) of non-variable stars, and thus are independent of the time series science data to be corrected; (2) the database covers a large area of the "Sweet Spot, so the methods do not require positional redundancy in the science data; (3) machine learning techniques in general allow for flexibility in training with multiple, sometimes unorthodox, variables, including those that trace PSF smear. We focus in this report on the K-Nearest Neighbors with Kernel Regression technique. (Additional communications are in preparation describing