From fragments to function: data-driven design of high-performance non-fullerene acceptors for organic photovoltaics

Abstract

The rapid advancement of organic photovoltaics (OPVs) depends critically on discovering non-fullerene acceptors (NFAs) with finely balanced optoelectronic properties. Yet, identifying optimal NFAs remains challenging due to the vast chemical space and complex, non-linear structure–property relationships. Here, we present a computational framework integrating physics-guided molecular fragmentation, hierarchical clustering, and combinatorial assembly with uncertainty-aware machine learning to accelerate NFA design. Beginning with 257 experimentally reported NFAs, we assembled a synthetically viable library of 500 000 NFAs with acceptor–donor–acceptor (ADA) structures. An evidential message-passing neural network (MPNN) was trained to predict oscillator strength (f), LUMO offset (ΔE_LUMO), absorption maximum (λ_max), and exciton binding energy (E_b), achieving high accuracy with built-in uncertainty quantification. Compared to the training distribution, our pipeline produced a deterministic enrichment of candidates with tightly converged, target-optimized values (f ≥ 1.5, ΔE_LUMO < 0.25 eV, E_b < 0.32 eV), in line with state-of-the-art OPV performance benchmarks. Quantum chemical validation confirmed prediction fidelity, with all deviations within 22%. This unified and interpretable framework provides a scalable route for rational NFA discovery and establishes a generalizable benchmark for machine learning-guided materials design in organic electronics.