Shear measurement bias II: a fast machine learning calibration method

We present a new shear calibration method based on machine learning. The method estimates the individual shear responses of the objects from the combination of several measured properties on the images using supervised learning. The supervised learning uses the true individual shear responses obtained from copies of the image simulations with different shear values. On simulated GREAT3 data, we obtain a residual bias after the calibration compatible with $0$ and beyond Euclid requirements for SNR>20 within $\sim 15$ CPU hours of training using only $\sim 10^5$ objects. This efficient machine learning approach can use a smaller data set since the method avoids the contribution from shape noise. Also, the low dimensionality of the input data leads to simple neural network architectures. We compare it to the state-of-the-art method Metacalibration, showing similar performances, and their different methodologies and systematics suggest them to be very good complementary methods. Our method could therefore be fastly applied to any survey such as Euclid or LSST, with less than a million images to simulate to learn the calibration function.