In order for AI systems to make the right decisions, they need to accurately "understand" the complex reality of the world around them. Directly learning models from observed data is often really hard or even impossible. Instead we need to take a modular approach of independently learning partial models of many different aspects of the world and putting them together as needed. Simulations -- not only of the partial Models but also including the sensors of the AI system -- then allows for generating exactly the synthetic data required for specific use cases. This can involve training of AI systems but also benchmarking, validation and -- most important for industry -- certification of "Trusted-AI" systems that provide guarantees about their functionality. But these "Trusted-AI“ systems also need to understand what they do not understand about their environment with the goal of improving their digital models about reality. In this talk, I will describe "Digital Reality", a comprehensive approach developed at DFKI that combines modeling, simulation, AI and high-performance computing in an unique and integrated way.