Permeability prediction using Support vector machines

Abstract

This thesis explores the possibility of calculating the permeability of materials with regression based of classification software instead of using famous physics formulae, like the Kozeny-Carman equation. The reason for this is that regarding permeability, no universal and exact formula has been discovered to date; the existing formulae depend on the constitution of the materials. The chosen model is based on classification from support vector machines. This classification algorithm was chosen because support vector machines have a history of showing accuracy comparable to those of other methods in recognizing various data and because they have a rigorous mathematical base. The thesis consists of a theoretical part and an applied one, where the first describe the basis on which the results rely and the second explains how certain parameters are calculated and used for the classification algorithm to perform well. It is shown that the classification algorithm surpasses the famous Kozeny-Carman equation in terms of accuracy of the calculations for fibre structures. It is also shown that it suffices to extract parameters from two dimensional images of the three dimensional structures to gain equal precision as if the whole three dimensional structure is taken into account. This raises hope that microscopy images can be used to calculate the permeability of materials. Finally it is shown that the content of the training set is more important than its size for the support vector machine to perform well.