Kriging and Neural Network models for pressure losses across perforated plates
summary
The gist
In this paper, two novel data-driven models based on kriging and neural networks (NN) are proposed to predict pressure losses across perforated plates with circular perforations in turbulent flows.
In short
Two data-driven models, kriging and neural networks, were developed to predict pressure losses across perforated plates in turbulent flows using experimental data from various sources. The models consistently outperformed existing empirical formulas for most plate configurations, showing strong predictive capability for practical CFD applications.
Key concepts
- Kriging Predictor
- A statistical method used to estimate unknown values based on known data points. It calculates a predicted pressure loss ($ψy(x^*)$) and its associated error (mean-squared error) by fitting a spatial distribution to the experimental measurements, providing a smooth prediction surface.
- Multilayer Perceptron (MLP)
- A type of artificial neural network architecture used as a surrogate model. It consists of an input layer with two variables, one or more hidden layers for complex feature extraction, and an output layer that predicts the single target variable, trained using the Adam optimization algorithm.
- Darcy-Forchheimer Equation
- An empirical formula describing pressure loss across perforated plates in turbulent flow. It accounts for both viscous drag (Darcy term) and inertial drag (Forchheimer term), which depends on fluid velocity, plate characteristics like porosity ($ε), and thickness ($δ/D).
- Porosity ($ε$)
- A key plate characteristic representing the fraction of the cross-sectional area that is open (perforated) rather than solid. The paper found that pressure loss predictions are highly sensitive to this porosity, especially at low values.
Terminology used across episodes
This episode discusses
- Kriging and Neural Network models for pressure losses across perforated plates · Paper Radio
- Jacobian-Enhanced Neural Networks
- A fluid flow model for the pressure loss through perforated plates
The paper
Kriging and Neural Network models for pressure losses across perforated plates · Read on arXiv
Shuai Lia
Chalmers University of Technology
DOI: 10.1016/j.tsep.2026.104978
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Kriging and Neural Network models for pressure losses across perforated plates".
Mira: Here's the gist of the paper. In this paper, two novel data-driven models based on kriging and neural networks (NN) are proposed to predict pressure losses across perforated plates with circular perforations in turbulent flows.
Kai: The predictive performance of the proposed models is assessed and compared against widely used empirical formulae.
Mira: It is found that the proposed models consistently outperform existing empirical models for most perforated plate configurations contained in the experimental datasets.
Kai: Model Development and Data Sources The two data-driven models are developed using two sets of experimental data available in the literature.
Paper summary: Kai: So we're looking at this paper, "Kriging and Neural Network models for pressure losses across perforated plates," and what they claim is that they developed two new data-driven models using kriging and neural networks to predict pressure losses in turbulent flows through plates with circular holes.
Mira: It sounds like the main point here is that these models can actually outperform the old empirical formulas we usually rely on for this kind of problem, which is pretty interesting because those old formulas often don't capture all the complicated interactions between plate shape and flow conditions.
Lev: From an engineering standpoint, if these new models are accurate, it means we might be able to make predictions faster or more reliably when we can't afford a full computational fluid dynamics simulation for every single design variation.
Kai: They developed these using two separate sets of experimental data from Yavuzkurt and Catchen and Méry and Sebbane, which is a pragmatic choice since there's just so little experimental data available in the literature on this topic.
Mira: That partitioning of the data is important because it lets them train their models without using the exact same measurements for testing, which keeps the performance assessment unbiased against those specific datasets.
Lev: When you're building something for real hardware, you always have to worry about how well it generalizes, and training on limited datasets means we need to be really careful about what those models can actually handle outside of that training set.
Kai: The kriging predictor uses an equation where the prediction at a specific point is based on a mean plus a weighted sum of the actual measurements scaled by some matrix inverses, and they also provide an error calculation for that kriging predictor.
Mira: And then they pair that up with a multilayer perceptron, or MLP, which is basically their neural network architecture, using full-batch Adam optimization to train it on the flow data.
Lev: For us building real hardware, the complexity of training these deep networks means we have to consider how much computational overhead that adds when you try to deploy this prediction on a system with limited memory or processing power.
Paper summary: Kai: They compared their results against the Bae and Kim model, which only considered the Forchheimer component of the pressure drop, and found that both their kriging and neural network models generally gave more accurate predictions for most plate setups in the experimental data.
Mira: The paper notes that for the kriging model, they saw no noticeable errors at all, while the neural network model showed only minor deviations across most of those plates.
Lev: Minor deviations are okay if you're dealing with real-world uncertainty, but we need to see how those predictions hold up when we introduce other variables like different fluid properties or slight geometric imperfections that aren't in the training data.
Kai: They also tested these models by feeding them into numerical simulations, where they introduced a pressure drop source term into the momentum equations as negative delta p over delta zero when the airflow approaches at zero incidence angle.
Mira: It's interesting because their RANS predictions matched their model predictions quite well, which confirms that these data-driven approaches are suitable for actual computational fluid dynamics applications and aren't just theoretical exercises.
Lev: And they also found that the number of streamwise cells they used for modeling the perforated plate didn't really change the solutions, which suggests these models are somewhat robust to how you discretize the geometry in a simulation.
Kai: So, overall, even though the kriging model showed slightly better agreement with the experimental data than their neural network model, both approaches demonstrated a strong predictive capability for pressure losses across perforated plates.
Mira: The authors point out that the response surfaces of both models show that the predicted pressure losses are quite sensitive to porosity when it's low, where those surfaces become markedly steeper.
Lev: That sensitivity to porosity at low values is something we need to watch closely when designing things, because if you get a small error in measuring or modeling the plate's actual porosity, that error gets magnified significantly in the prediction.
Kai: However, they do have limitations, specifically mentioning that although these models work for turbulent flows over a wide range of Reynolds numbers, they are limited to specific ranges of plate porosity and the thickness to diameter ratio.
Mira: So you can't just use these models for every single engineering problem without checking if your specific configuration falls within those tested ranges for porosity and thickness.
Paper summary: Lev: That makes sense; you need constraints on the input parameters, otherwise, you're just trusting a curve that might break outside its tested domain.
Kai: The paper also provides an appendix with some empirical formulae for pressure loss across perforated plates, like the Darcy-Forchheimer equation and how it relates to normalized pressure drop based on terms like delta p over rho U zero times two D times K epsilon Re p plus alpha delta d.
Mira: It’s useful context because it shows where their new data-driven approach fits in relation to the established theoretical framework, showing that their results align with what we expect from those classical models under certain conditions.
Lev: The way they expressed the Darcy drag term, showing how it scales with D squared K plus epsilon alpha D Re p, is a concrete way to see exactly what physical contributions are being modeled by these terms in the pressure loss equation.
Kai: It seems like these data-driven models offer a feasible framework for modeling pressure losses across perforated plates because they consistently outperform existing empirical models for most configurations in the experimental datasets used.
Mira: The implication is that we can use machine learning to fill in those gaps where traditional physics equations are hard to apply directly, especially when dealing with complex geometries and turbulent conditions.
Lev: For real hardware development, this means we could potentially use these models for rapid initial design iterations before committing to expensive high-fidelity CFD runs for every single variant.
Kai: It’s a promising avenue because it shows that using kriging and neural networks can give us good predictive performance without requiring the massive computational resources that traditional solvers demand.
Mira: The paper suggests that these models are applicable to a wide variety of engineering problems because they cover many practical configurations of porosity and thickness.
Lev: That's the key, but we still have to manage the input data quality carefully; if the training data is noisy, even a good model will produce noisy predictions.
Kai: So, while both kriging and neural network models show strong predictive capability, we should keep an eye on those limitations regarding porosity ranges when we apply this work to specific design tasks.
Conclusion: Kai: So this paper is about using kriging and neural networks to predict how much pressure loss happens when fluid flows through plates that have circular holes, specifically in turbulent flows.
Mira: Right, it’s looking at how these two data-driven models stack up against the old, traditional formulas for calculating that pressure drop.
Lev: The authors used experimental data from Yavuzkurt and Catchen and Méry and Sebbane to train these models because they didn't have much literature on this stuff before.
Kai: They trained a kriging model, which is like using existing measurements to predict new ones based on spatial relationships, and a neural network called an MLP for the same thing.
Mira: And the results show that both of these new models actually beat the older empirical models for most of the plate setups they tested.
Lev: That’s interesting because it means we might get better estimates without needing a full computational fluid dynamics simulation every single time we want to check a design.
Kai: Yeah, but they also highlighted that the predictions are really sensitive to things like porosity when it gets low, which is a big caveat for any real-world application.
Mira: Exactly, so we have these good predictive tools, but we gotta be careful about the input parameters like how open those holes are.
Lev: And they did test these models against actual numerical simulations where they added a pressure drop term into the momentum equations, and the results matched up quite well with their model predictions.
Kai: That’s good because it confirms that these data-driven approaches aren't just theoretical exercises; they actually work in practice for modeling fluid flow.
Mira: It suggests that machine learning methods can be a practical way to predict these kinds of complex physical phenomena when the underlying physics are too messy to model directly.
Lev: And since they have limitations on the range of porosity and thickness ratio, we still need to respect those boundaries when trying to use these models for new hardware designs.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T