Special CMX Seminar

With the invention of digital computers in the mid 1940's, John von Neumann envisioned numerical weather prediction (NWP) as a revolutionising application of this technology and famously stated in a talk at Princeton in 1950: "All processes that are stable we shall predict. All processes that are unstable we shall control." In today's language, von Neumann had sown the seeds for the modern concept of NWP as a digital twin (DT) of actual weather processes in the atmosphere, its physical twin (PT). Since the days of von Neumann, DTs have found ground breaking applications in science and engineering. Still many challenges remain when considering highly complex multi-scale processes arising from, e.g., aeronautics, cognition, ecology, and pharmacology. Mathematically speaking, a DT can often be described as a partially observed Markov decision process (POMDP). Solving POMDPs computationally constitutes one of the most challenging problems around. Still, tremendous progress has been made in closely related fields such as data assimilation, uncertainty quantification, control and optimisation, and model reduction. A key emerging questions is thus how we can successfully synthesise these advances into a tool broadly applicable to DT. In my talk I will attempt to map out one possible roadway.