Robust DNN Surrogate Models with Uncertainty Quantification via Adversarial Training