Adaptive PINN Framework Balances Physics and Data for Heat Transfer Prediction

A novel self-supervised physics-informed neural network (PINN) framework has been developed to effectively balance physics-driven and data-oriented supervision for scientific machine learning in situations with limited data. Unlike previous PINNs that employed fixed or heuristic weights for physics residuals and data loss, this new method features a learnable blending neuron that adjusts the contributions of each component based on their uncertainties. This innovation allows for stable training and better generalization without the need for manual adjustments. Additionally, the framework incorporates a transfer learning strategy that repurposes representations from related fields to adapt to new physical systems with scarce data. Validation for predicting heat transfer in liquid-metal miniature heat sinks used only 87 CFD data points, resulting in an error rate below 8%, surpassing traditional neural networks and kernel methods. This research is detailed in arXiv preprint 2605.05217.