Adaptive surrogate models in heterogeneous computer environments

High-fidelity simulation models provide insights into complex systems involving high computational effort. Intelligent surrogate models can reuse knowledge gained from former simulation results. Hence, cheap yet accurate surrogates enable accessible model insights anywhere by approximating and accelerating the expensive models.

A main goal of this project is to implement holistic surrogate models of dynamic bio- and structural dynamical systems on weak hardware by significantly reducing computational costs. For this purpose, we develop novel and adaptive surrogate modeling techniques that combine nonlinear dimensionality reduction and coordinate identification with approximations of latent dynamics to save resources. Clustering strategies and error estimators can be used to identify time and space domains of interest; based on good provenance data, we select optimal surrogate models based on the current demand and domains. In addition, those surrogate models can incorporate sensor measurement feedback to synchronize the virtual models with the real world and improve their accuracy and robustness.