MODEL BUILDING USING GENETIC ALGORITM FOR DATA FITTING Manuel Borregales Jose Cappelletto Miguel Asuaje manuelantoniobr@gmail.com cappelletto@usb.ve asuajem@usb.ve Universidad Sim´ on Bol´ ıvar, Sartenejas , Baruta, Caracas-Venezuela Andrea Shmueli Milan Stanko andrea.shmueli@ntnu.no milan.stanko@ntnu.no Norwegian University of Science and Technology, 7491, Trondheim, Norway Abstract. There are many methods for data fitting. Each of them try to find the best coeficients configurations to minimizea residual function. Those coefficientscan be obtained from a linear or non linear model. If the model for a data set data is known, finding the coefficients to minimize the residual function is simple. However in some engineering problems the fitting model for a data set is unkown. In this work, a methodology using genetic algorithms is proposes to find a model or correlation that better fit a set of data. To accomplish this implementation, binary trees and Pr¨ ufer encoding are used. This methodology allows to find a model or correlation regardless the number of independent variables to be used, at a low computational cost and it does not require previous information on the behavior of the data. Two applications from this methodology will be shown. The first one is about an axial force on mud pump check valve. The hydrodynamic actuating force is fitted as a function of the flow and separation from the valve seat. The second is about the prediction of liquid droplet entrainment in stratified/annular flow. The liquid droplet flow was estimated as a function of three dimensionless variables. Data set for each application were divided in two groups, a group to build the model and another to validate it. A model to estimate axial force with 6% RMS error for the check valve problem and a model to predict liquid droplet flow with 16% RMS error for the two- phase flow problem were obtained. Key words: Genetic Algorithms, Data Fitting, Two-Phase flow, Droplet entrainment, Flow in valves