Developing new chemical systems, like functional materials and sustainable catalysts, is inherently a multi-scale problem. Interactions between atoms and molecules (nanometers) determine properties on a meter scale. While traditional simulation techniques currently reach their boundaries, introducing data science, artificial intelligence (AI), and machine-learning (ML) approaches into the fields of computational chemistry and theoretical materials science led to significant progress in the last years. For example, one of the applicants has developed atomic descriptors, conversions from coordinates of the participating atoms to a feature vector usable by AI algorithms. These are crucial to ensure that the AI model obeys physical constraints, like the conservation of energy and momentum. However, current research software for AI in, e.g., computational chemistry, exhibits severe shortcomings with respect to classical software engineering metrics – sacrificing maintainability, reproducibility, reusability, or sustainability for efficiency and time-to-application-specific-results. Furthermore, there is a lack of best practice examples targeted to researchers from the application disciplines. The current project aims at narrowing this gap by providing: a) an improved application software based on AI for materials science applications w.r.t. scientific capabilities, efficiency, and software design; b) generalized software tools for AI in data-scarce / data-generating applications from science and engineering; c) best practice examples from materials science targeted to scientists; and d) teaching materials targeted to doctoral students from science and engineering, introducing them to typical scientific workflows for AI with TensorFlow. Within PN3, our project obviously addresses RQ 3 (machine-learned models). It also contributes to RQ 1 (Scale bridging for particle models) by allowing to transfer the accuracy from ab-initio training data to molecular dynamics simulations. Within the whole of SimTech, our project deals with FC1 (multi-X methods) via its contributions to materials science. Especially the molecular dynamics side of the project addresses FC3 (bridging data-poor and data-rich regimes) by exploiting a limited number of ab-initio training data to obtain physical properties of ionic liquids in different length and time scales.