Cross-laboratory generalization failure in perovskite solar cell machine learning: A diagnostic protocol and evaluation threshold

초록

Machine learning (ML) models for perovskite solar cell power conversion efficiency are commonly evaluated using within-laboratory cross-validated R2, with high values often interpreted as evidence that composition–efficiency relationships have been learned. Here, we show that this metric can diverge sharply from cross-laboratory predictive performance on the same dataset. Using a controlled three-laboratory cohort of 819 p-i-n perovskite solar cells and five model families spanning linear, regularized, and ensemble approaches, within-laboratory R2 values of up to 0.44 collapse to leave-one-laboratory-out (LOLO) R2 values as low as −3.05. Cross-laboratory prediction errors are several times larger than those obtained by predicting the mean, and the collapse persists across all five model classes. Restricting the feature matrix to composition descriptors only partially mitigates the failure (LOLO Random Forest (RF) R2 improves from −3.05 to −2.10), indicating that the problem is not a feature-engineering artifact but a structural property of multi-laboratory datasets in which composition, device architecture, and laboratory identity are intrinsically confounded. To systematically expose and evaluate this failure mode, we introduce a diagnostic framework combining within-laboratory, full-dataset, and LOLO cross-validation with SHapley Additive exPlanations stability analysis and analysis-of-variance decomposition. Under this framework, meaningful cross-laboratory generalization requires LOLO R2 > 0; the composition-only LOLO RF R2 of −2.10 reported here indicates that current methods fail to satisfy this criterion in the present cohort. Similar dissociation is likely across materials ML studies trained on pooled multi-laboratory datasets. Lab-aware evaluation should therefore be a standard component of ML reporting, not an optional addition.

키워드