We aim at establishing performance portable implementations for high performance fluid simulations and developing large scale data analyses for extreme-scale simulations. For high performance computing, we have demonstrated that a performance portable implementation in C++ alone is possible without harming the readability and productivity. As a data-driven studies, we have developed two deep learning models. Firstly, we have developed a surrogate model to predict the plume dispersion in a complicated urban area for the emergence response capability to contaminant gas leakage events. Second, we have developed a deep learning based Sub-Grid-Scale model which allows the large eddy simulation with 1/10 of grid points compared to direct numerical simulations.

• A Deep learning model for plume concentration prediction in the realistic urban area
• Examples of performance portable implementations of mini-applications in Kokkos, OpenMP, Thrust and stdpar