Study in the Multiscale Materials Characterization group

Abstract

Gunpowder is an ancient multipurpose invention from 9th century in China. The chemical composition and physical structure of gunpowder have been developed for centuries to fulfill many purposes which lead into the development of smokeless powder or propellants. Knowledge about the chemistry surrounding propellant is vastly studied and known but precise study of the interior structure of individual propellant grains are lacking. This study validates X-ray microtomography as a method to analyse the inner physical structures of propellant grains from 18 different batches. I measured two grains from each batch. My focus of interest was on interior material domain identification, their sizes, porosity, pore sizes and the correlation of these with the recipe of the propellant. To aid my analysis I used a Python script and batch specific recipes for almost all of the batches provided by Nammo Vihtavuori Oy. Particles identified from the images are nitrocellulose, potassium bitartrate, potassium sulfate, surface finishing compound and two metals: tin and bismuth. Identified components do not change their original grain like shape while inside propellant and their size is also preserved. However, variations inside the batches are high, resulting in different sized particles and pores in varying locations for each grain. Potassium nitrate salt had the only notable correlation to the inner structure of propellant grains: the more there was salt the higher the porosity and thinner the walls between the pores are. Unfortunately, large variations in the atomic number of component elements introduced artifacts into the images, and high variations between grains prevented me from gathering statistically relevant data. Nevertheless, imaging of propellant with X-ray tomography (X-CT) has proven useful giving lots of previously unknown information of the physical structure of propellant grains and their composition, which can be used to develop new improvements to propellant. Method could also be improved on to gain more precise information about the propellant, for example reducing the image artifacts.

Abstract

Methods for nowcasting lightning using weather radar data were developed using machine learning models. Reflectivity was selected as the main feature for the prediction. The purpose was to examine if machine learning applications could be used to nowcast thunderstorms with minimal data sets. The emphasis was to find out a model which is based on binary image classification and doesn’t require large sets of training data to work sufficiently. Convolutional neural network was the first choice. Accuracy for the model was 0.83. Another approach was made using random forest model. Precision for class 0 (no lightning) was 0.52, and for class (recorded lightning) 1, 0.90 and with total accuracy of 0.88. To improve the sets more features should be used and possibly larger data sets.