Abstract. AI is increasingly penetrating the production industry. Today, however, AI is still used in a limited way in a production environment, often focusing on a single production step and using out-of-the-box AI algorithms. AI models that use information spanning a complete production line and even larger parts of the product lifecycle could add significant value for production companies. One of the barriers for creating such AI models, however, is the fact that they typically require information access over a wide range of heterogeneous data sources, including time series, (physics) models, (non-)relational databases, data gathered through web APIs, etc. This is especially relevant considering the trend towards high-mix-low-volume production, where the amount of measurement data for each product variant is typically low and AI models can be significantly improved by adding information from as many additional data sources as can be found, including data from PLM and MES systems and even less tangible data sources, such as operator experience. In this paper, we suggest a digital twin architecture to support the complete AI lifecycle (discovering correlations, learning, deploying and validating), based on a knowledge graph that centralizes all information. We show how this digital twin could ease the information access challenge and pose opportunity for a wider application of AI in production industry. We illustrate this approach using a simplified industrial example of a compressor housing production.