Analysis and Type Identification of Ancient Glass Relics Based on Machine Learning

The chemical composition of glass is a special kind of compositional data. For the statistical analysis of compositional data, the classical statistical methods are no longer applicable due to the limitation of the fixed sum and the compositional data usually have many missing values. Therefore, the composition data are first changed logarithmically and then interpolated by a suitable method, and the modified EM algorithm has the best interpolation effect among several interpolation methods. The effect of weathering on the glass surface on the chemical composition data was investigated by applying ANOVA to the interpolated data. The results showed that for the same batch of glass artifacts, the type of glass had a significant effect on weathering or not, and further obtained that glass weathering had a significant effect on the changes of chemical composition such as silica, potassium oxide, and lead oxide. Then using gray correlation analysis to study the correlation between the chemical compositions of different types of glass artifacts, we obtained that among the high potassium type of glass, the highest correlation between silica and sodium oxide was 0.883, and among the lead-barium type of glass, the highest correlation between barium oxide and silica was 0.91. For the classification of glass types, this paper used the emerging machine learning methods to classify, in which SVM, KNN, and random forest methods for ten-fold cross-validation all have a correct rate of 100%, indicating that machine learning has good predictive effect on the classification of glass types. Also comparing with traditional glass classification methods and cluster analysis methods, it is found that using machine learning methods for glass type classification can also get good results.