A Machine-Learning-Based Surrogate Model for Internal Flow Nusselt Number and Friction Factor in Various Channel Cross Sections

Abstract

The design of a wide variety of cold plates, heat sinks, and heat exchangers relies on geometry-specific models and empirical correlations for Nusselt number and friction factor to select the heat transfer surface in an iterative manner. Machine learning techniques are being used across many branches of science to develop more generalized surrogate models that can predict such transport properties. To collapse the catalogue of available heat transfer surfaces into a single predictive tool, the present study develops machine-learning-based surrogate models for the Nusselt number and friction factor values for fully developed internal flow. The Nusselt number and friction factor data for the available constant cross section flow geometries in literature are compiled, pre-processed, and then used to train the models. To reduce variance and avoid overfitting, an ensemble method (bootstrap aggregating) is used, with the base learners as three-layer artificial neural networks (ANNs). The number of layers and nodes in each layer of the ANN is determined based on a grid search across a wide range of hyperparameters. The predictions from the surrogate model are compared to the training data set of ~94,000 points across both Nusselt number and friction factor, considering angular, aspect ratio, and rotational variations of 35 distinct channel cross sections. The trained model has a mean percentage error in the Nusselt number and friction factor of ~2.0% and ~0.7%, respectively, for the entire dataset.