Improving Physical Parametrizations in Climate Models using Machine Learning