SuPi - Super-resolution for physics-informed neural networks

• Applicant: Dr.-Ing. Michael Mommert
• Subject Area: Fluid Mechanics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 566613562

Heat transport is a quantity that is either minimised or maximised in many engineering problems. For this reason, convective heat fluxes are of particular interest in the study of thermal convective flows, from fundamental research to practical applications. Until recently, the fields of the heat fluxes could only be measured by combining measurement techniques such as particle tracking velocimetry and particle image thermometry, which introduced a significant complication to the measurement setup. However, recent research has shown that machine learning approaches in the form of physics-informed neural networks (PINNs) have the potential to extract temperature information from velocity fields in buoyancy-driven flows. However, this capability so far only extends to moderately turbulent flows when fully resolved velocity fields are provided. The proposed project aims to extend the framework for reconstructing temperature fields with the goal to enable the PINN to generate structures on smaller scales than those present in the provided data, which has so far prevented the application of PINNs to highly turbulent flows. Thus, the proposed extension targets the major obstacle of the handling of velocity data that does not fully resolve the complete spectrum of scales. Coincidentally, the project also addresses the current limitations of PINNs stemming from their multi-objective optimisation.

• DFG Programme: Research Grants