Machine learning prediction of perovskite solar cell efficiency across standardized experimental data

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Perovskite solar cells combine high efficiencies with solution-based, low-temperature processing, yet translating laboratory gains into stable, scalable devices remains difficult. To address this, we compiled and standardized ∼40,000 device entries encompassing n-i-p and p-i-n stacks, transport layers, absorber composition and thickness, and back contacts. Each record includes forward and reverse J-V scans to capture hysteresis effects. Across the corpus, power-conversion efficiency is governed primarily by short-circuit current density and fill factor; open-circuit voltage shows a smaller contribution, and layer-thickness variables display weak linear trends, pointing to strongly non-linear device behavior. Six regressors, Linear Regression, SVM, Random Forest, SGD, Bayesian Ridge, and XGBoost were trained on 10 k, 20 k, and full datasets for both scan directions. Ensemble models performed best: Random Forest achieved the lowest errors and, together with XGBoost, reached R 2 ≈ 0.97–0.98 across conditions. Aggregated trends indicate practical levers for device design: mixed-cation, lead-based absorbers with band gaps near 1.5–1.7 eV, intermediate ETL/HTL thicknesses (ETL: ∼ 30–40 nm and HTL: 150–200 nm), and moderate annealing temperatures and durations (Perovskite: ∼100–150 °C for ∼10–60 min, ETL: ∼150 °C for ∼30–60 s, HTL: ∼100–120 °C) are typical of high-performing stacks, while challenges remain for scalable deposition and durable, low-cost contacts. The benchmark offers validated prediction baselines and concrete guidance for process optimization.

Links and sources

Comments