Generalization-Oriented Cross-Literature Process Informatics Reveals Transferable but Context-Dependent Fabrication Signals in Perovskite Solar Cells

Abstract

Literature-scale databases provide an opportunity to learn fabrication–performance relationships in perovskite solar cells, but conventional device-level validation can confound transferable process information with publication-specific experimental structure. Here, we develop a generalization-oriented process-informatics framework using a provenance-controlled cohort of 41,015 manually extracted devices from 7292 digital object identifier-indexed publications in the NOMAD-hosted Perovskite Solar Cells Database. Material/device descriptors, process-reporting indicators, and actual fabrication variables were analyzed separately using Extreme Gradient Boosting, Light Gradient Boosting Machine, and random forest models under random-row, publication-held-out, and rolling future-year validation. Under random splitting, approximately 97% of test records shared publication identity with training data, substantially inflating apparent predictive performance. Nevertheless, after complete digital object identifier (DOI) separation, adding fabrication descriptors reduced mean absolute prediction error by approximately 0.19–0.20 percentage point in power conversion efficiency (PCE) across all three model families (an incremental gain, not the absolute model error of approximately 3.2 PCE percentage points). The process-related predictive gain was reproduced in two well-supported future-year tests and persisted in complete-case analyses. Among physically defined fabrication-process blocks, solvent-recipe descriptors showed the most reproducible cross-publication and temporal transfer. Their incremental predictive value remained after expanded control for material, formulation, measurement, reporting, and other process variables and after restriction to single-step spin coating. However, exploratory within-publication N,N-dimethylformamide–dimethyl sulfoxide comparisons involving only 22 DOI groups provided insufficient evidence for a common monotonic mixing-ratio response or shared optimum. These results distinguish transferable predictive information from a universal process law and suggest the use of literature-scale process informatics to identify promising process–context relationships for subsequent experimental validation, rather than to prescribe a universal solvent ratio.