Navigating high-dimensional fabrication-parameter space in organic photovoltaics device optimization by a multi-tier machine learning framework

Organic photovoltaics (OPV) have achieved significant advances over recent decades, driven by synergistic innovations in molecular design and device engineering. However, precise morphological control within bulk-heterojunction active layers remains a crucial barrier to commercial viability, primarily due to the high-dimensional parameter spaces and complex interdependencies among processing variables. To overcome this challenge, we established a standardized materials-processing-performance database integrating donor/acceptor pairs, nine key active layer processing parameters, and device efficiencies. This database, curated from over a decade of experimental results, resolves critical data heterogeneity issues and provides the field's most comprehensive optimization resource. Leveraging this resource, we developed a novel three-tiered machine learning framework employing gradient boosting regression trees to progressively decode active layer processing complexities. Our strategy initiates with single-parameter models for targeted optimization, advances through stage-combined models revealing intra-process synergies (e.g., solvent-additive interplay), and culminates in a global optimization tier. Remarkably, this final tier achieves unprecedented performance, demonstrating >0.9 overall Pearson correlations, and >80% success rates in identifying optimal nine-dimensional configurations. Experimental validation on 78 novel systems confirms exceptional generalization, yielding >75% accuracy in predicting either optimal or secondary parameters across eight active layer processing conditions. This work establishes a robust framework for navigating processing complexity in high-dimensional spaces, enabling accelerated optimization of OPV photoactive layers and providing a transferable data-driven paradigm for rational process design in emerging photovoltaic technologies.