Coordinate-Free Scientific Machine Learning Reveals Sequence-Dependent Electronic Regimes in Peri-Metalated Polyacenes: From Dominant Size/Composition Trends to Reproducible Local Arrangement Effects

Abstract

Rigid carbon frameworks in organic semiconductors restrict the tunability of their electronic and spin properties. Peri-metalation of polyacenes with coinage metals, particularly gold and copper, offers a route to electronic regimes not attainable in conventional organic systems, yet navigating this hybrid space often requires exhaustive quantum chemical sampling. This study integrates density functional theory with a small-data scientific machine learning framework to show that global electronic trends, including bandgaps, ionization energies, and electron affinities, can be captured by minimalist, coordinate-free descriptors. The hybrid architecture combines an analytical baseline defined only by inverse ring size and Au/Cu counts with coordinate-free residual learning that utilizes discrete metal sequence and topology descriptors. The original 53-descriptor residual model offers a chemically comprehensive representation, whereas a reduced 4-descriptor model assesses the persistence of principal predictive trends following significant dimensionality reduction. By circumventing explicit atomic coordinates, geometric parameters, orbital energies, wavefunctions, and interaction energies as model inputs, both models successfully recover chemically meaningful electronic properties and trends across the polyacene series while remaining sensitive to subtle local sequence effects. Systematic model–DFT deviations serve as diagnostic indicators, revealing that Cu-rich extended acenes represent a regime where the learned size and composition scaling is quantitatively insufficient, thereby necessitating further electronic structure analysis. While gold metalation yields stable, predictable electronic structures, copper incorporation drives the system into “emergent” regimes characterized by extreme bandgap narrowing and near-degenerate singlet–triplet states. This work establishes a framework in which machine learning performance itself marks the boundary of simple chemical trends, offering a rational approach to the discovery of low-bandgap, spin-sensitive hybrid semiconductors.