Work Package 3

In each of the use-cases, the various experts collaborate closely to integrate all necessary aspects in obtaining Exascale-ready AI solutions for real-world problems. The overarching goals, which mirror to the use-cases, are

• the development of in-situ feature detection techniques, i.e., Convolutional Neural Networs are trained in-situ to support automatic detection of important physical features in simulations to gain further insight into complex phenomena,

• solving control and optimization problems, i.e., to have error-controlled extrapolation methods at hand,

• perform Physics-Informed Deep Learning to use physical constraints as priors in the loss of Machine / Deep Learning algorithms to restrict learning to a space of physically reasonable / sound solutions,

• replace expensive physical processes by AI models and surrogates to make simulations computationally more efficient,

• to develop new numerical models with the help of AI, i.e., to derive at new models for multi- physics and/or multi-scale problems,

• and to uncover non-obvious correlations of physical phenomena by means of AI methods to (i) increase computational performance and (ii) increase the accuracy of modern numerical simulation methods.