EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Robust Multi-Output Prediction of Perovskite Solar Cell Parameters via Multi-Task Learning

Khaled Chahine, Mohamad Arnaout, Marc Al Atem, Abdallah El Ghaly, Hassan N. Noura

Inventions · 10.06.2026 · 11 (3) · 59

DOI: 10.3390/inventions11030059 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Inventions, online erschienen am 10.06.2026. Verfasst von Khaled Chahine und 4 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Conventional machine learning models for perovskite solar cells predict photovoltaic parameters independently, disregarding the physical constraint PCE=Voc×Jsc×FF/100. This approach can yield mutually incompatible predictions for the four parameters, a failure mode that has not been hitherto quantified in the perovskite solar cell literature. This paper proposes a multi-head neural network with a shared backbone, physics-guided feature construction, and task-specific prediction heads, and validates it on 7176 SCAPS-1D simulations across 12 perovskite compositions. When benchmarked against architecturally matched single-task baselines, the multi-task model, optimized via 5-fold cross-validation, achieves R2 values of at least 0.994 for all four targets, with cross-fold standard deviations of 0.001. In particular, fill factor prediction improves from R2=0.617±0.254 (single-task) to 0.994±0.001 (multi-task), a 233-fold reduction in cross-fold standard deviation. Application of a physical consistency metric developed in this work reveals that 36.5% of single-task predictions exceed a 2 PCE-unit implausibility threshold, compared to only 0.01% for the multi-task model. The multi-task model outperforms the single-task baseline in all 20-fold target comparisons, with large effect sizes (Cohen’s d=1.33–8.93). These results confirm multi-task learning as an effective approach for achieving robust, stable, and internally consistent predictions in simulation-based photovoltaic virtual screening.

Zitieren

APA

Chahine, K., Arnaout, M., Atem, M. A., Ghaly, A. E., & Noura, H. N. (2026). Robust Multi-Output Prediction of Perovskite Solar Cell Parameters via Multi-Task Learning. Inventions, 11(3), 59. https://doi.org/10.3390/inventions11030059

Vancouver

Chahine K, Arnaout M, Atem MA, Ghaly AE, Noura HN. Robust Multi-Output Prediction of Perovskite Solar Cell Parameters via Multi-Task Learning. Inventions. 2026;11(3):59. doi:10.3390/inventions11030059

BibTeX
@article{Chahine2026,
  author = {Chahine, Khaled and Arnaout, Mohamad and Atem, Marc Al and Ghaly, Abdallah El and Noura, Hassan N.},
  title = {Robust Multi-Output Prediction of Perovskite Solar Cell Parameters via Multi-Task Learning},
  journal = {Inventions},
  year = {2026},
  volume = {11},
  number = {3},
  pages = {59},
  doi = {10.3390/inventions11030059}
}
Kostenlos starten →