Predicting Notebook Heat Exchanger Performance Using a Neural Network Approach

Thermal engineers must design the heat exchanger geometry of an actively cooled notebook computer to meet an overall thermal resistance target for their thermal solution. Geometric parameters for the heat exchanger must be chosen to meet system, blower and thermal constrains. The typical approach is for the thermal engineer to estimate the adequate heat exchanger geometry and to iterate the design using feedback from correlations and simulations. These feedback mechanisms have trade-offs between accuracy and time often resulting in long iteration cycles to arrive at an optimal design. In this paper a neural network approach is utilized to predict heat exchanger air-flow impedance and thermal resistance using a large CFD generated training dataset. A 3-level 8-factor full factorial DOE on notebook representative heat exchanger configurations was created and solved using IcePak resulting in 3^8=6,561 distinct runs. This dataset was then used in MATLAB to train a neural network for both air-flow impedance and thermal resistance with resulting R correlation coefficients greater than 0.99. The result is an accurate and fast method for the thermal engineer to iterate the heat exchanger geometry for optimal performance. Also demonstrated in this paper is the applicability and effectiveness of using neural networks for multi-factor thermal predictions.