We present a novel scientific document discoverysystem inspired by molecular chemistry and AI-driven drug discovery. Our approach treats document tokens as atomic units, which are combined to form "molecular" representations ofmathematical documents. We employ a probabilistic framework to maximize the likelihood of forming coherent mathematicaldocuments while minimizing the probability of random token combinations and non-STEM document tokens. To achieve this, we develop a token embedding scheme that maps property vectors to a musical keyboard, effectively representing each token as a musical chord. We further differentiate between STEM and non-STEM documents by introducing a harmonic constraint on adjacent nodes in document graphs. Specifically, STEM documents are characterized by polyphonic harmonization of adjacent node vectors, whereas non-STEM documents exhibit dissonant relationships. Our system integrates a graph neural network/transformer decoder architecture, trained end-to-end to generate STEM documents from input graphs. This innovative approach has the potential to revolutionize scientific document discovery and retrieval.