Danger Zone: Establishing Buffers for Enhanced Classification in BPT Diagrams

This study utilizes unsupervised machine learning, specifically the uniform manifold approximation and projection (UMAP) algorithm, to classify optical spectra originating from star-forming regions, Seyferts, and low-ionization (nuclear) emission-line regions (LI(N)ERs) based on their line ratios. Typically, the ionization source of a region is determined from intensity ratio of different combinations of pairs of spectral lines. However, using current boundary definitions, $\sim10$\% of spectra change classes between diagnostic diagrams. We apply the machine learning technique to $\sim$1.3 million optical spectra from 6,439 galaxies observed in the MaNGA survey. By training UMAP on consistently classified data, we can classify these ``ambiguous'' spectra, and delineate boundary zones where such ambiguities arise. Furthermore, we identify physically interesting subsets within the ambiguous spectra. Future work will incorporate additional parameters, such as alternative emission line ratios and velocity dispersions, to enhance classification accuracy.