Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift
- 1Department of Industrial Chemistry, Faculty of Applied Physical Sciences
- 2Enugu State University of Science and Technology (ESUT
Abstract
Machine-learning models for organic photovoltaics (OPVs) can achieve strong performance under random train-test splits, but such evaluation may overstate prospective generalization when literaturederived datasets contain repeated publications and material systems. Here, two complementary OPV benchmarks were constructed to separate interpolation from chemical extrapolation: a provenanceclean OPV-DB cohort containing 21,590 experimentally reported devices and a processing-rich Wen-Zhang-Ma cohort containing 994 unique physical processing conditions. Random, publication-heldout, donor-acceptor (D:A)-pair-held-out, acceptor-held-out, and donor-held-out validation regimes were compared, together with structure-only, processing-only, and combined structure-processing representations. Extra Trees performance on OPV-DB declined from R2 = 0.813 under random validation to R2 = 0.362 under donor holdout. In the processing-rich benchmark, adding processing information reduced MAE by 19.5% under random validation but by only 1.0-2.4% under chemically grouped validation. This effect was strongly associated with material-system familiarity: processing reduced MAE by 27.1% for test records whose D:A pair was represented in training, compared with 1.8% for completely unseen pairs. Ensemble dispersion was useful for ranking errors under random validation (Spearman rho = 0.462) but lost reliability under donor holdout (rho = 0.047), where the most uncertain 20% of predictions captured only 16.1% of the largest errors. A cross-dataset transfer audit further showed that a common unmodified RDKit representation covered 73.54% of the processing benchmark and only 38.02% of records with PCE >= 15%, creating substantial chemicalselection bias. These results show that validation design is central to claims of OPV model reliability, that processing information primarily refines predictions within familiar chemistry, and that both uncertainty estimation and transfer benchmarking require explicit domain-shift and representation audits.
Related
- Prediction of High-Performance Donor–Acceptor Pairs for Organic Photovoltaics with Machine Learning
- Beyond the Ligand Applicability Domain: Generalization and Local Resolution in Structure-Aware Foundation Models for Affinity Prediction
- Chemical Generalization Boundaries in Fire-Extinction Prediction from a Public Surfactant Formulation Database
- Publication-block calibration reveals wide uncertainty in PLGA release endpoint prediction
- Thermodynamically Constrained Deep Learning for Hydrogen Sorption Prediction: Overcoming Classical Isotherm Generalization Failure through Physics-Informed Cross-Lithology Transferability
BibTeX
@misc{orji2026reliable,
title = {{Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift}},
author = {DENNIS OBINNA ORJI},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008466/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008466/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift
AU - DENNIS OBINNA ORJI
PY - 2026
DA - 2026/09/08
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008466/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008466/v1
AB - Machine-learning models for organic photovoltaics (OPVs) can achieve strong performance under random train-test splits, but such evaluation may overstate prospective generalization when literaturederived datasets contain repeated publications and material systems. Here, two complementary OPV benchmarks were constructed to separate interpolation from chemical extrapolation: a provenanceclean OPV-DB cohort containing 21,590 experimentally reported devices and a processing-rich Wen-Zhang-Ma cohort containing 994 unique physical processing conditions. Random, publication-heldout, donor-acceptor (D:A)-pair-held-out, acceptor-held-out, and donor-held-out validation regimes were compared, together with structure-only, processing-only, and combined structure-processing representations. Extra Trees performance on OPV-DB declined from R2 = 0.813 under random validation to R2 = 0.362 under donor holdout. In the processing-rich benchmark, adding processing information reduced MAE by 19.5% under random validation but by only 1.0-2.4% under chemically grouped validation. This effect was strongly associated with material-system familiarity: processing reduced MAE by 27.1% for test records whose D:A pair was represented in training, compared with 1.8% for completely unseen pairs. Ensemble dispersion was useful for ranking errors under random validation (Spearman rho = 0.462) but lost reliability under donor holdout (rho = 0.047), where the most uncertain 20% of predictions captured only 16.1% of the largest errors. A cross-dataset transfer audit further showed that a common unmodified RDKit representation covered 73.54% of the processing benchmark and only 38.02% of records with PCE >= 15%, creating substantial chemicalselection bias. These results show that validation design is central to claims of OPV model reliability, that processing information primarily refines predictions within familiar chemistry, and that both uncertainty estimation and transfer benchmarking require explicit domain-shift and representation audits.
ER -