The objective of this talk is to identify the challenges and develop a unique and comprehensive setting of data-knowledge environment in the realization of the development of ML models. We review some existing directions including concepts arising under the name of physics informed ML. Key ways of elicitation and accommodation of domain knowledge are investigated. An impact on the structuralization of the ML architectures and the ensuing implications on the interpretability, explainability and credibility as well as semantic stability are studied. We investigate the representative topologies of ML models identifying data and knowledge functional modules and interactions among them. The detailed considerations on the facet of explainability including new ideas of semantic stability are covered. We also elaborate on the central role of information granularity in this area.