Heat transfer correlations by symbolic regression

We describe a methodology that uses symbolic regression to extract correlations from heat transfer measurements by searching for both the form of the correlation equation and the constants in it that enable the closest fit to experimental data. For this purpose we use genetic programming modified by a penalty procedure to prevent large correlation functions. The advantage of using this technique is that no initial assumption on the form of the correlation is needed. The procedure is tested using two sets of published experimental data, one for a compact heat exchanger and the other for liquid flow in a circular pipe. In both situations, predictive errors from correlations found from symbolic regression are smaller than their published counterparts. A parametric analysis of the penalty function is also carried out.