Renana Gershoni-Poranne

Delocalization is one of the foundational principles of modern chemical theory. From resonance structures to band theory, it provides the framework for how we understand conjugated systems. Yet delocalization also carries an implicit assumption: that a molecule’s electronic behavior is an emergent, holistic property—not reducible to the sum of its parts. And yet… what if it is? In this talk, I explore the surprising success of additivity in predicting magnetic properties of conjugated molecules. Through a series of studies on NICS-XY-Scans,1,2 bond currents, and current densities,3 we show that even large polycyclic systems can be modeled accurately using only a small set of mono-, bi-, and tricyclic subunits. This is not merely a computational convenience—it is also a window into the behavior of polycyclic aromatic systems. Strikingly, this pattern of locality also emerges from an entirely different direction: deep learning. Using graph neural networks trained only on small molecules, we achieve accurate predictions of magnetic behavior in systems many times larger.4 That these models succeed reinforces a central point: delocalization, at least in its magnetic manifestation, behaves as though it is fundamentally local. This rethinking opens new avenues for how we understand, predict, and ultimately design π-conjugated systems—not by ignoring delocalization, but by redefining its scope.