Simulation supports various fields of design and engineering in creating better solutions. Especially, to create a sustainable built environment, complex interdependencies of engineering need to be considered during early design phases. This engineering integration requires the analysis of a high number of variants in early design phases. Lack of information, modelling effort and computation times prohibit such extensive use of simulation. Recent research promotes the use of data-driven models created by machine learning. These models form surrogates to replace physical simulation, reduce effort of modelling and computation drastically, and link to the available information in form of designoriented variables. However, such data-driven models have shortcomings in terms of generalizability outside training data and explainability due to their black-box

character. The talk will present a novel component-based method that overcomes these limitations, makes data-driven models reusable as components in a broad range of novel cases, and that allows for internal insights in engineering quantities within data-driven models. First, the generalization and flexibility integrated in digital modelling will be demonstrated. Second, the benefits of explainability and interpretability by observing internal engineering quantities will be illustrated by energy efficient building design.