Feature selection in machine learning algorithms for nuclear DFT

We present a new process to identify features for machine learning algorithms that will be used to make predictions in nuclear structure calculations. We focus on the particular task of quantifying the model bias in nuclear binding energies calculated with Density Functional Theory (DFT). While the model bias can be directly calculated when experimental data is available, only an estimate can be made in the absence of measured data. Although machine learning algorithms have found applications in many areas of nuclear science, some of these applications require to extrapolate the input variables to regions outside of the domain that was used to train the algorithm. Current approaches quickly lose predictive power when the input variables are extrapolated. Our process of feature selection avoids this situation by incorporating the distribution of different features in all regions of the nuclear chart. This allows for reliable predictions of the model bias and a direct improvement of DFT calculations.