Since the first applications of Deep Learning to a classification problem in, where a substantial improvement with respect to standard methods efficiency has been shown, Deep Learning approaches are widely spreading to improve the performance of selections. An important aspect to consider is that a model based on many input variables usually produces results more difficult to explain and understand. In this sense DL networks are exemplary even more than trees or support vector machines: understanding how input and output are causally connected appears often very arduous. After a few layers, the representation of predictions in terms of input variable features is barely comprehensible in terms of human logic. Deep Learning interpretability is often referred to as “explainable Artificial Intelligence problem”. Explainability is particularly relevant in scientific experiments, as it helps to detect bias, to strengthen against potential adversarial perturbations and to guarantee that only meaningful variables affect the output, preserving causality in the model reasoning.  We work to develop new teqniques to improve the explainability of models for classification problems in HEP experiments.