A Systematic Approach for Solving Large-Scale Problems by Neural Network: Open Refrigerated Display Cases and Droplet Evaporation Problems

Abstract

A systematic approach for solving a large-scale design problem is proposed. The method consists of building a multidimensional test case matrix connecting an output (scalar or vector) to an input vector. Every element of the input vector spans over its possible minimum and maximum values with one or more levels in between. The experimental and validated computational methods are combined to find the output as a function of all permutations of the input variables. The results are used to train a neural network program to perform the proper interpolation for any other expected scenario. The number of required experiments is obtained asymptotically by adding more data to the neural network training set and examining the error. The applicability and feasibility of this approach is shown in two different problems: first, predicting and minimizing the infiltration rate into an open refrigerated display case; and second, predicting the evaporation rate of sessile microdroplets on a nonpermeable surface.