Estimating spatially varying convective heat transfer coefficients (HTCs) is a challenging inverse problem in thermal engineering, particularly for jet impingement cooling, where the coefficient changes across the cooled surface. A new arXiv preprint investigates a physics-informed neural network (PINN) approach that combines generalized boundary representations with transfer learning to tackle this estimation task.
PINNs embed physical governing equations into the neural network's loss function, allowing the network to infer unknown parameters from data. The study specifically looks at how generalized boundary representations can better capture the spatial variation of HTCs, and how transfer learning can reuse a trained model for new but related cases.
The abstract does not report numerical results or comparisons with other methods, so the effectiveness of the proposed combination is not yet quantified. The contribution at this stage is the methodological framework: using generalized boundary representations and transfer learning together in a PINN-based inverse estimation pipeline for jet impingement cooling.