Atomistic Simulation and Symbolic Regression Reveal Design Rules Linking Blend Morphology to Photovoltaic Performance in PM6:NFA Solar Cells

The performance of non-fullerene-acceptor (NFA) organic solar cells is governed by the nanoscale donor–acceptor morphology of the active layer, yet quantitative design rules linking molecular packing to device metrics remain limited. Here, we integrate fully atomistic molecular dynamics simulations with symbolic-regression-based machine learning to derive interpretable morphology–performance relationships for PM6:NFA blends. Ten representative morphologies were characterized using geometric, packing, and interaction descriptors capturing density, orientational order, interfacial spacing, aggregation, and free volume. After variance inflation factor–based descriptor screening, SISSO (sure independence screening and sparsifying operator) symbolic regression was employed to construct compact analytic models for short-circuit current ( J SC ), open-circuit voltage ( V OC ), fill factor (FF), and power conversion efficiency (PCE max ). The resulting low-dimensional expressions accurately reproduce experimental performance trends ( r = 0.97–1.00) while retaining physical interpretability. Three primary morphology axes emerge: (i) a density–anisotropy–free-volume axis governing percolation and charge collection, (ii) an orientation-mismatch axis influencing interfacial charge separation and transport directionality, and (iii) a clustering/compactness axis associated with energetic disorder and voltage losses. These results translate complex atomistic morphology into actionable, physically grounded design principles for polymer:NFA solar cells and provide a generalizable framework for morphology-driven optimization of next-generation organic photovoltaic materials.