Optimization of hydrogen acoustic microfluidic pump based on loss function-enhanced physics-informed neural networks

The coupling design of acoustic streaming effects and sharp-edged array structures within a microfluidic pump is essential for enhancing hydrogen transport performance. This study introduces an enhanced physics-informed neural network (PINN) that combines multiple loss functions, including gradients and partial differential equation (PDE) residual losses. A detailed analysis of the impact on both nonlinear fitting accuracy and optimization efficiency is provided to achieve efficient optimization of the acoustic field, flow field, and structure. The results indicate that enhancing the PDE residual loss in PINNs contributes most significantly to improving fitting accuracy, establishing it as the central component in Bayesian hyperparameter optimization. A notable interaction is observed between the PDE residual loss and the Sobolev gradient loss, indicating that a well-balanced combination of different loss terms can substantially enhance the accuracy of surrogate models. The enhanced PINN reduces the sample size required for fitting a nine-dimensional Ackley function by 86% compared to a conventional PINN while maintaining high fitting accuracy with an R2 value of 0.99. Although different optimization algorithms yielded consistent optimal results, the enhanced PINN reduced the number of iterations by 41.2% compared to the standard PINN, exhibiting superior stability, robustness, and faster convergence toward the optimal solution. The optimized asymmetric sharp-edged array structure increases hydrogen flow by 68.3%, minimizes the impact of vortices on the main flow, and generates a stronger driving force through an unobstructed Rayleigh–Schlichting flow. These findings provide critical engineering insights for optimizing hydrogen transport designs.