Density Functional Theory (DFT) is a leading computational framework for modeling nuclear properties in unmeasured or unmeasurable nuclei. However, it is an incomplete theory with conceptual and predictive shortcomings that result in complicated biases in relation to the experimental findings. Artificial neural networks can flexibly learn a mapping to these bias distributions, but their ability to generalize beyond the available data is severely restricted. Recent efforts that use machine learning to better model nuclear properties have been inadequately tested for their predictive power far outside the training domain. This thesis is a continuation of current research that provides a superior template for learning these biases. After reserving a large out-of-sample region of the nuclear chart for testing, the neural network takes DFT-calculated moments and quantities as features. This work measures the effect of filtering, or screening, the included input features based on the overlap their probability distribution shares with the testing region. The results show dramatic gains in testing performance compared to neural networks trained on only atomic and mass numbers or all available DFT features, and indicate a reliable method to improve nuclear DFT predictions on unmeasured nuclei.