Physics-Guided Inverse Design of Nonfullerene Acceptors via a Deep-Learning-Accelerated Genetic Algorithm

The discovery of high-performing nonfullerene acceptors (NFAs) for organic solar cells (OSCs) is challenged by the vastness of chemical space and the need to satisfy multiple tightly coupled electronic criteria. Here, we present a physics-informed generative framework that integrates an evidential message-passing neural network (MPNN) with a constraint-encoded genetic algorithm (GA) to enable inverse molecular design guided directly by quantum-relevant descriptors. Instead of relying on empirical OSC efficiency surrogates, the GA optimizes three key molecular-level properties known to govern charge generation efficiency─oscillator strength (f), exciton binding energy (E_b), and the LUMO-LUMO+1 energy gap (ΔE_LUMO)─while enforcing structural validity and chemical realism throughout evolution. The combined MPNN-GA workflow efficiently explores a diverse chemical landscape and converges toward synthetically plausible NFAs that satisfy stringent multiobjective constraints. Predicted properties show strong agreement with quantum chemical benchmarks, confirming the reliability of the surrogate model. Pareto analyses further reveal that the generative pipeline captures established quantum-chemical trade-offs and extends the accessible design frontier by identifying candidates that simultaneously exhibit high f, low E_b, and suppressed ΔE_LUMO. These results demonstrate a scalable and interpretable approach for physics-driven inverse design of next-generation NFAs, offering a generalizable strategy for molecular discovery in organic electronics.