Kriging and neural network models for pressure losses across perforated plates

Predicting pressure losses across perforated plates is critical for engineering applications like heat exchangers, combustion chambers, and noise reduction. Traditional empirical models are often complex and struggle to maintain accuracy across diverse configurations. This paper investigates whether data-driven approaches—specifically kriging and neural networks—can provide a more accurate and generalizable alternative for modeling these pressure losses in turbulent flow regimes.

To build the models, the author utilized two distinct sets of experimental data from the literature, covering various perforated plate configurations (varying pore diameter, spacing, and plate thickness). The dataset was partitioned into training and test sets to ensure unbiased validation. The kriging model was formulated using a Gaussian process approach, while the neural network model employed a multilayer perceptron architecture optimized via Leave-One-Out Cross Validation (LOOCV). To demonstrate practical utility, the author implemented these models as momentum source terms within Reynolds-averaged Navier-Stokes (RANS) simulations of two-dimensional channel flows.

Both the kriging and neural network models demonstrated superior predictive performance compared to widely used empirical formulas across the test dataset. The response surfaces generated by these models revealed that pressure loss is highly sensitive to porosity, particularly at low values, while showing relatively weak sensitivity to the plate thickness ratio. Furthermore, the RANS simulations confirmed that these data-driven models can be seamlessly integrated into computational fluid dynamics (CFD) workflows, yielding results in excellent agreement with experimental measurements. This suggests that data-driven surrogates are a robust and feasible framework for complex fluid dynamics problems where experimental data is scarce.