Interpretable machine learning reveals morphology-mediated performance pathways in organic solar cells

Abstract

Organic solar cell (OSC) performance is governed by coupled electronic, morphological, optical, and transport processes; however, the relative contribution of intrinsic electronic descriptors and processing-controlled structural parameters remains difficult to quantify in heterogeneous experimental datasets. In this work, interpretable machine learning is used as an applied device-physics tool to identify morphology-mediated performance pathways in OSCs. Rather than treating power conversion efficiency (PCE) only as a direct prediction target, the photovoltaic sub-metrics open-circuit voltage \left(V_{oc}\right), short-circuit current density \left(J_{sc}\right), fill factor (FF) are first predicted and then used to reconstruct PCE through the standard photovoltaic relation. Structure-only, electronic-only, and combined descriptor models are compared to evaluate the relative predictive roles of morphology-related processing variables and intrinsic electronic descriptors. Model interpretation is performed using permutation feature importance, SHAP analysis, and partial dependence/ICE curves. The results show that structural and processing descriptors provide stronger predictive information than electronic-only descriptors for the measured photovoltaic outputs. Structure-only models achieved test R^2values of approximately 0.730 for V_{oc}, 0.662 for J_{sc}, and 0.726 for FF. Active-layer thickness emerges as the dominant structural descriptor, particularly for J_{sc}and FF, reflecting the competing effects of photon absorption, carrier-transport distance, recombination, and charge extraction. SHAP and partial-dependence analyses reveal nonlinear and context-dependent thickness effects modulated by solvent, additive, annealing, and device architecture. Reconstructed PCE shows partial agreement with experimental PCE, confirming that structural descriptors mediate efficiency through photovoltaic sub-metrics while electronic and interfacial losses still constrain performance. This study demonstrates that morphology-aware interpretable machine learning can provide physics-based insight into OSC structure–processing–performance relationships and guide device optimization.