Graph-based Construction of the Astrochemical Molecular Inventory toward Sgr B2(N)

Molecular inventories are essential for exploring chemical networks and physical conditions, yet their systematic analysis is often hindered by severe line confusion and heterogeneous excitation conditions in spectral surveys. In this paper, we propose a graph-based framework for molecular inventory analysis in Sgr B2(N) that integrates six complementary descriptors into a heterogeneous molecular graph, including source size, rotational temperature, line width, velocity offset, vibrational state, and a Self-Referencing Embedded Strings-based molecular structure representation. Molecular embeddings are learned via multiscale, metapath-guided random walks that jointly encode chemical structure and astrophysical co-occurrence, enabling relationally consistent inference of molecular column densities across 20 subregions of Sgr B2(N) with R² values ranging from 0.64–0.91. Furthermore, we introduce an interpretability framework combining attribution analysis, prototype retrieval, contrastive molecular analysis, and an embedding-based abundance-informed prioritization metric, showing that the learned embedding preserves molecular structural similarity while capturing environmental drivers of abundance variation. By unifying prediction, interpretation, and abundance-informed prioritization within a single pipeline, this work provides a scalable, data-driven approach to astrochemical inventory analysis in spectral surveys.