A Learning-Based Optimal Uncertainty Quantification Method and Its Application to Ballistic Impact Problems