Session: 15-02 Numerical Studies of Internal Cooling 1

Paper Number: 154276

Neural Network Modeling for Predicting Local Heat Transfer in Jet Impingement: A Comparison With Traditional Correlations 

Various authors have reported piecewise correlations in the literature to capture local heat transfer characteristics in jet impingement for different jet-to-target plate spacings and Reynolds numbers. Developing a single, comprehensive correlation using traditional data-fitting approaches is challenging. Therefore, a neural network is trained using the experimental data of Katti and Prabhu (2008) to generate a unified predictor for local heat transfer performance. The neural network architecture consists of four layers with 40, 32, 16, and 8 neurons, respectively. The input variables include the Reynolds number, jet-to-target plate spacing, and radial distance from the stagnation point, while the output variable is the local Nusselt number. The hidden layers utilize the ReLU (Rectified Linear Unit) activation function, and the output layer uses a linear activation function. To optimize the model, the Adam (Adaptive Moment Estimation) optimizer is employed with a constant learning rate of 0.01. The model is trained using 80% of the dataset, while the remaining 20% is reserved for validation. The validation dataset plays a critical role in preventing overfitting and ensuring that the model generalizes well to unseen data.

The accuracy of the neural network is evaluated by comparing it to the correlation reported by Katti and Prabhu (2008). Key performance metrics such as the regression coefficient (R²), mean absolute error (MAE), and root mean square error (RMSE) are used to quantify the predictive performance. For the traditional correlation, the R², MAE, and RMSE are 0.985, 3.35, and 4.88, respectively. In contrast, the neural network model yields significantly improved results with an R² of 0.998, an MAE of 1.25, and an RMSE of 1.88. This demonstrates that the neural network not only successfully learned from the experimental data but also predicted the local heat transfer characteristics with greater accuracy than the traditional correlation. By leveraging the ability of neural networks to model non-linear relationships, researchers and engineers can develop more robust predictors for a wide range of operating conditions.

Presenting Author: Dylan Olson Saint Louis University

Presenting Author Biography: Dylan Olson is a senior undergraduate student in the Department of Aerospace and Mechanical Engineering at Saint Louis University.

Authors:

susheel singh Saint Louis University
Prashant Singh The University of Tennessee, Knoxville
Dylan Olson Saint Louis University