This contribution presents a methodological framework for formalising the specifications of a simulator based on deep machine learning. The aim of this research, which combines modelling epistemology and theoretical computer science, is to develop verification and validation techniques for critical systems requiring the predictive power of machine learning modelling.