TLDR: A new study presents a CNN-based surrogate model that accurately predicts heat transfer (Nusselt number) in impinging jet arrays with dynamic inlet/outlet configurations in real-time. Trained on CFD simulations, the model achieves high accuracy (NMAE < 2%) and can extrapolate predictions to higher Reynolds numbers, enabling efficient model-based temperature control for complex thermal management systems.
In the realm of advanced thermal management, controlling temperature precisely across surfaces is crucial for various applications, from manufacturing parts and cooling gas turbines to managing electronic systems. Impinging jet arrays, which involve multiple jets directing fluid onto a surface, are highly effective for this purpose because they can create localized areas of high heat transfer. However, the complexity arises when these jet systems need dynamic control, where each jet can independently act as an an inlet or an outlet, or even be shut off, leading to a vast number of possible flow configurations.
Traditionally, predicting heat transfer in such complex systems relies on computational fluid dynamics (CFD) simulations. While CFD offers high accuracy, its significant computational cost makes it impractical for real-time applications, such as model-based temperature control. Imagine needing to adjust jet configurations instantly to maintain a specific temperature; CFD simply isn’t fast enough.
To overcome this limitation, a recent study introduces a novel solution: a surrogate model based on a Convolutional Neural Network (CNN). This model is designed to predict the Nusselt number distribution – a key indicator of heat transfer efficiency – in enclosed impinging jet arrays in real time. The researchers, including Mikael Vaillant, Victor Oliveira Ferreira, and Bruno Blais, developed this model by training it with data generated from implicit large eddy CFD simulations, specifically for Reynolds numbers below 2,000.
The study focused on two distinct jet array configurations: a five-by-one array and a three-by-three array, developing a separate surrogate model for each. The five-by-one model was trained with data from 83 simulations, while the three-by-three model used 100 simulations. A significant innovation presented in this work is a method to extrapolate these predictions to higher Reynolds numbers (up to 10,000) using a correlation-based scaling technique, making the model applicable to a wider range of industrial conditions.
The performance of these surrogate models is impressive. On validation data, the five-by-one model achieved a normalized mean average error (NMAE) below 2%, and the three-by-three model performed even better with an NMAE of 0.6%. This high accuracy means the models can reliably predict heat transfer patterns. Experimental validation further confirmed the models’ predictive capabilities, showing good agreement with real-world temperature measurements.
Also Read:
- Advancing Fluid Dynamics: A New Benchmark and Deep Learning Model for Particle Image Velocimetry
- Enhancing Time-Dependent PDE Predictions with Deep Ensembles
This research lays a vital groundwork for developing model-based control strategies in advanced thermal management. By providing real-time predictions of heat transfer, the surrogate model enables dynamic modulation of flow rates and jet arrangements, which is essential for precise temperature control in complex systems where traditional methods fall short. For more details, you can refer to the full research paper available at this link.
