Physics-informed neural network simulation of conjugate heat transfer in manifold microchannel heat sinks for high-power IGBT cooling

This study explores the application of Physics-Informed Neural Networks (PINNs) in modeling fluid flow and heat transfer dynamics within intricate geometric configurations, focusing on manifold microchannel (MMC) heat sinks designed for efficient high-power IGBT cooling. A deep neural network architecture comprising two sub-PINNs, one for flow dynamics and another for thermal behavior, is developed, each initialized with a sine activation function to capture high-order derivatives and address the vanishing gradient problem. Comparisons between PINN and CFD simulations reveal close agreement, with both methods showing an increase in pressure drop and a decrease in temperatures as inlet velocity increases. Discrepancies arise in scenarios with rapid flow pattern or gradient changes, highlighting PINNs' sensitivity to geometric complexity and numerical stability. Overall, this study underscores PINNs' potential as a promising tool for advancing thermal management strategies across various engineering applications. • Physics-Informed Neural Networks (PINNs) as an innovative approach to model fluid flow and heat transfer in complex geometries is introduced. • Close agreement between PINN and CFD simulations, affirming PINNs' accuracy, bolstering their credibility is achieved. • PINNs' sensitivity to geometric intricacies and numerical stability, particularly in abrupt flow pattern or gradient changes is identified.