Extraction of component bases from mixed spectra using non-negative matrix factorization with dissimilarity regularization

Abstract

We propose a method of automatic pattern decomposition for mixed spectra, based on non-negative matrix factorization. This method uses a regularization term that increases volume coordinated by the decomposed bases. This treatment enhances dissimilarities among the bases and is suitable for expressing the natural component bases of spectra, which generally differ from each other. This regularization term shows better accuracy in a test using virtual spectrum datasets compared to conventional regularizations. We also test the proposed method with a high-throughput X-ray diffraction dataset measured in a synchrotron radiation facility. Our method can decompose the dataset into component bases, which are correctly assigned to background and chemical compounds, while results obtained by other regularization terms cannot sufficiently separate background peaks from and meaningful peaks. We also find that a combination of dissimilarity regularization with smoothness regularization can detect small meaningful peaks. These results show the advantage of dissimilarity regularization for the purpose of automatic spectrum decompositions.