TOWARDS UNDERSTANDING BIG DATA MOLECULAR SIMULATION USING MACHINE LEARNING ALGORITHMS
MetadataShow full item record
Analytical experiments and computer simulations have shown unusual thermal conductivity behavior of soft material under mechanical strain. At first, the conductivity decreases with the increase of the mechanical strain. After reaching a certain minima, with the application of additional strain, it starts increasing. While this behavior is confirmed via experiments as well as computer simulations using molecular modeling techniques, what causes this behavior is still considered an open question. This thesis studies the molecular simulations data from multiple simulations and investigates the use of machine learning techniques to understand the unusual thermal conductivity behavior of soft materials under mechanical strain. The data sets consists of terabytes of data with thousands of spatial as well temporal attributes. In particular, the influence of attributes related to molecular structure including bond length, bond angle, and dihedrals has been probed. Several machine learning and data mining techniques including clustering and association rule mining have been applied on different datasets with different thermal conductivity to find the relationship between the soft material molecular structure and thermal conductivity behavior. The results show that in the alkyls molecules the most important features that affect the thermal conductivity are the O-CH2 bond properties, especially when thermal conductivity is equal to 1.5±ε . Keywords: Thermal conductivity, bond angle, bond length, dihedral angle, association rule mining, ANOVA significance test, K-means clustering.