SizeResGraph-Guided GFN2-xTB Screening of Low-Gap Thiophene-Containing Polymer Repeat Units

Abstract

Abstract Low-gap conjugated polymers are widely studied for optoelectronic applications, but their vast repeat-unit space makes exhaustive quantum-chemical screening impractical. We developed SizeResGraph, a directed-bond graph neural network with a bounded size-dependent readout, to rank thiophene-containing repeat units by their hydrogen-capped GFN2-xTB HOMO–LUMO gaps. The model learned from 9,905 labeled structures curated from 540,650 public records. Across 10 scaffold partitions and five training seeds, SizeResGraph matched Chemprop D-MPNN within a prespecified ±0.010 eV RMSE equivalence margin. An end-to-end run ranked 419,564 prospective structures in 500.9 s using one RTX 4090 and 16 CPU threads. Direct GFN2-xTB calculations confirmed strong enrichment of the low-gap tail, with Low 5% hit fractions of 88.6% for SizeResGraph and 81.8% for equal-budget Chemprop. In a prediction-independent systematic sample, SizeResGraph predictions showed strong rank agreement with B3LYP-D4/def2-SVP gaps (ρ = 0.836). SizeResGraph combined with targeted quantum-chemical verification therefore provides a scalable route from large repeat-unit libraries to tractable sets of low-gap candidates.