Page 29 - Read Online

Page 22 of 34 Wang et al. Energy Mater. 2026, 6, 600064

Figure 5. Machine learning model prediction and experimental photovoltaic performance of perovskite solar cells. (A) The fitting graph of
PCE results by the CatBoost-based algorithm model, where red represents the training set and blue represents the test set. Reprinted with
permission [162] . Copyright 2024, John Wiley and Sons. (B) The distribution of experimental results and data points from the database in the
fitting graph. Reprinted with permission [162] . Copyright 2024, John Wiley and Sons. (C) J-V curves of devices with different perovskite
components via a two-step spin-spin sequential deposition method. Reprinted with permission [162] . Copyright 2024, John Wiley and Sons.
(D) Statistics of the PCE of PSCs with different groups based on 7 devices. The central line represents the median, the box limits
correspond to the upper and lower quartiles, and the whiskers extend to the minimum and maximum values. Reprinted with permission [162] .
Copyright 2024, John Wiley and Sons. (E) J-V curves of the regular devices and their champion photovoltaic performance. Reprinted with
permission [162] . Copyright 2024, John Wiley and Sons. (F) J-V curves of the inverted devices and their champion photovoltaic performance.
Reprinted with permission [162] . Copyright 2024, John Wiley and Sons.

most effective safeguard against overfitting remains rigorous out-of-sample evaluations. To validate the
generalization capability of this methodology, the team conducted a blind test on 17 newly designed
photovoltaic molecules. The predictive results demonstrated strong precision, achieving a mean absolute
error (MAE) of only 1.02% across the test group. Notably, the PCE of 10 novel molecular candidates were
predicted within a 1% error margin. This validation confirms that the model has captured the underlying
structure-property relationships rather than merely interpolating within known historical data, thereby
enabling rapid in silico evaluation of unexplored chemical spaces and providing a transparent and rational
blueprint for the design of future functional materials.

However, the most pervasive challenge in ML-based models arises when the available experimental datasets
are extremely limited. To address this small-sample bottleneck, the Co-Pilot for Perovskite Additive Screener
(Co-PAS) framework was developed around a fundamental shift in data partitioning and feature
representation. Pu et al. [165] abandoned traditional data augmentation strategies and instead focused on the
physical logic of data partitioning and the transferability of dimensionality reduction techniques. To address
the severe performance distortion and high variance caused by the random splitting of small datasets, they