Model calibration of the liquid mercury spallation target using evolutionary neural networks and sparse polynomial expansions