The present study investigates the dependency of prediction accuracy of an artificial neural network (ANN) on the network architecture using 65 different neural networks from seven architecture patterns. The accuracy of the ANNs is compared based on their capability to predict heat transfer coefficients of air-cooled heat sinks operating in laminar flow. Scattered input data is used for training the networks to make the modelling more realistic and closer to practical applications. The input variables for the neural network are heat sink width, channel height, channel length, number of channels, fin thickness, and Reynolds number. The output is heat transfer coefficient. The training process for all ANNs is performed using ReLU as the activation function. The accuracy of the neural networks is evaluated by the root mean square error. It is found that the prediction accuracy of an ANN is strongly dictated by the optimization of the network architecture, which corresponds to the proper number of hidden layers and the number of neurons at each layer. The most accurate architecture in the present study predicts heat transfer coefficients of 60% and 86% of heat sinks within ±10% and ±20% of the true values, respectively. However, an ANN with an unoptimized architecture results in a substantially reduced accuracy such that it predicts heat transfer coefficients of only 19% and 30% of heat sinks within ±10% and ±20% of the true values, respectively.
In the present study, artificial neural network (ANN) models are developed to predict heat transfer coefficient (ℎ) and pressure drop (∆𝑃𝑃) in cold plates (CPs) with surface roughness operating in turbulent flow. Roughness sizes range from zero (smooth surface) to 0.5 mm, and Reynolds numbers vary from 3,170 to 10,560. The RNG 𝑘𝑘 − 𝜀𝜀 model is used to simulate turbulent flow. Input data for the ANN models are prepared by simulating three-dimensional steady state turbulent flow and heat transfer inside the CPs. Separate multilayer neural networks are selected to predict ℎ and ∆𝑃𝑃. Both ANN architectures include two hidden layers with 1,024 neurons in each layer. The accuracy of the training process and the neural network is assessed by the mean absolute error. Both ANN models show excellent predictions as the predicted ℎ and ∆𝑃𝑃 are within ±1.2% and ±2.6% of the simulated values, respectively. Since roughness is an inevitable consequence of additive manufacturing, the present study suggests that accurate ANN-based models can be used as promising design tools for optimizing additively manufactured CPs. While roughness improves heat transfer, it leads to a higher pressure drop. As a result, accurate ANN models can be used to design additively manufactured cooling systems with an optimized range of roughness to improve heat transfer while operating within the allowed pressure drop and pumping power.
Thermal performances of multi-layered cold plates (CPs) with varying numbers of channels are investigated through threedimensional simulation of laminar flow and heat transfer. Thermal performances are characterized by the maximum temperature and temperature variation across the heating surface. The thermal performances are presented as functions of flow rates and pumping power to provide better insight on CP's practical applications. It was found that at both a given flow rate and pumping power, increasing the number of layers monotonically enhances the heat transfer rate.; however, the percentage of enhancement of heat transfer is reduced by increasing the number of layers beyond two due to additional thermal resistance experienced between the lower-level channels/layers and the heat source. The findings suggest the existence of a threshold number of layers such that beyond that threshold, the heat transfer is not enhanced.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.