3.1.1.

Optimizing surface capping layer

The degradation of perovskite films directly impacts the lifetime of perovskite photovoltaic devices. As a result, surface treatment of perovskites to form a low-dimensional perovskite capping layer has emerged as an effective strategy to enhance the stability of these photovoltaic devices. Nevertheless, a significant challenge persists in identifying appropriate materials for surface modification. Recently, ML has garnered significant attention in the development of passivation materials for perovskite photovoltaics.^78–81 The primary emphasis has been on the prediction and design of materials for passivation. This involves compiling a database of passivation materials from literature and then employing ML for predictions, with the use of SHAP to interpret the importance of features. In addition, there is a significant trend in comparing the results of ML predictions with actual experimental outcomes in devices, as well as with insights from DFT calculations. This comparison serves to further substantiate the reliability of ML applications.

Hartono et al. developed an ML method to select overlay materials to suppress perovskite degradation.⁷⁸ They collected 21 organic salts from the PubChem database as overlay materials, using the structure and chemical characteristics of organic molecules as well as device fabrication conditions as model inputs, while using the initial degradation values and degradation rates of perovskites as model outputs [Fig. 2(a)]. Regression models [Fig. 2(b)] and SHAP value analysis [Fig. 2(c)] were employed to analyze the relationship between features and stability. It was found that a smaller number of hydrogen bond donors and a smaller topological polar surface area in organic molecules are associated with increased stability of the ${\mathrm{MAPbI}}_3$ films. The best-performing organic halide, phenethyltriethylammonium iodide (PTEAI), successfully extended the stability lifetime of ${\mathrm{MAPbI}}_3$ by $4\pm 2$ times compared to control ${\mathrm{MAPbI}}_3$. Moreover, characterization techniques (X-ray photoelectron spectroscopy and Fourier-transform infrared spectroscopy) have also revealed that PTEAI effectively forms a capping layer by modifying the surface structure and chemistry. This alteration correlates with the suppression of methylammonium loss and the formation of both ${\mathrm{PbI}}_2$ and oxygen-containing compounds at the surface of the perovskite, ultimately achieving the goal of protecting the perovskite from degradation.⁸²

Fig. 2

(a) Schematic illustration of design rules for capping layer of PSCs. (b) The cross-validated root mean square error of various machine-learning models. (c) The feature importance ranking obtained from the RF regression algorithm and SHAP library, showing the chemical properties and processing conditions in descending order of importance (rank). The yellow and purple color indicates high and low values of a given feature, respectively. Reproduced with permission from Ref. 78 (CC-BY).

Recently, Zhi et al. collected a dataset of 46 experimental data points documenting the improvement in power conversion efficiency after interface passivation with 19 different ammonium salts at various concentrations.⁷⁹ These ammonium salts encompassed 14 from the initial dataset and five additional data selected via Latin hypercube sampling (LHS). LHS achieves diversity and uniformity sampling by dividing the range of values for each parameter into equally spaced subintervals and ensuring that there is only one sample point within each subinterval. The model inputs were 12 molecular descriptor features and the concentration of the ammonium salt precursor solution. The model output was the rate of improvement in device PCE. It is noteworthy that, to enhance precision, they employed an ensemble learning method that integrates five distinct ML regression models with various architectures to train and make predictions on the dataset. Similar to the previous work, SHAP analysis was also utilized to investigate the impact of molecular features on device efficiency, identifying hydrogen bond donors, hydrogen atoms, and the molecular lipid−water partition coefficient (molecular LogP or MolLogP) as the three most critical molecular features for the passivating ammonium salts. Ultimately, they employed the trained ML model to screen 10 suitable ammonium salts from 112 candidates in the PubChem database and experimentally validated six different ammonium salts. Among them, 2-phenylprop-1-ammonium iodide achieved PCEs of 22.36% and 24.47% for FAMACs and FAMA-based PSCs, respectively.

3.1.2.

Degradation mechanism

Despite numerous strategies having been employed to enhance the stability and power conversion efficiency of perovskite photovoltaics, the factors affecting stability and PCE remain elusive. Hu et al. constructed an ML regression model to investigate the key factors influencing the efficiency and stability of PSCs (Fig. 3).⁸³ They selected 102 sets of PSC device data from published literature. Through feature engineering, they retained key parameters: Eg (bandgap), size (the grain size of perovskite films, inversely representing grain boundary density), Rn (surface roughness at the interface between the perovskite film and the HTL), Trap (trap density), and TRPL (average fluorescence lifetime from time-resolved photoluminescence measurements) as model inputs. These were utilized with the aim of predicting two critical outcomes: power conversion efficiency and stability. Several ML regression models containing support vector regression (SVR) and ridge regression were employed to discern the relationships between model inputs and outputs. Among these, SVR demonstrated the best performance. Moreover, it was found that the bandgap had the greatest impact on efficiency, while surface roughness and grain size predominantly influenced stability. Finally, guided by predictive insights, various annealing temperatures were strategically employed to modify grain sizes, and three unique organic molecules were utilized to change surface roughness. This approach yielded PSCs with an impressive 23.4% efficiency and exceptional long-term stability, maintaining 97.6% of their initial efficiency after an extended period of 3288 h under ambient conditions. It is noteworthy that the results from DFT calculations also show a remarkable correlation with the predictions of the ML model.

Fig. 3

Working principle of ML: relevance analysis—experimental verification—one with the best properties. Complex multivariable analysis by ML. PVK is perovskite. Reproduced with permission from Ref. 83 © 2022 Wiley‐VCH.

3.1.3.

Optimizing electron transport layer

An ETL that matches the energy levels of the perovskite layer and exhibits high electron mobility is fundamental for ensuring efficient electron injection and extraction. In addition, it can impact the density of defect states at interfaces between the perovskite and the ETL, thereby preventing detrimental interfacial recombination. As a result, an appropriate ETL is essential for the performance of perovskite photovoltaic devices. This indicates the need not only to choose suitable ETL materials but also to perform surface modification or doping of the selected layer, to ensure efficient electron transport, and to facilitate perovskite crystal growth.

She et al. utilized a two-step ML approach, involving classification and regression, to study suitable ETLs (Fig. 4).⁸⁰ Initially, they extracted 1820 performance data points of PSCs from 795 articles published between 2013 and 2020, forming the first dataset. The data were then categorized based on the materials of the ETLs, dividing them into groups, such as ${\mathrm{TiO}}_2$-based, ${\mathrm{SnO}}_2$-based, and other ETL-based groups. In this dataset, the ETLs were undoped. To prevent sample imbalance, the samples were further classified into three categories based on their efficiency: category A for efficiency below 9%, category B for efficiency ranging from 9% to 18%, and category C for efficiency above 18%. By selecting nine features that influence the power conversion efficiency through feature engineering, they used these as model inputs with power conversion efficiency as the output, training different ML classification models on this dataset. Ultimately, they further investigated feature importance based on the classification model, finding that ${\mathrm{TiO}}_2$ or ${\mathrm{SnO}}_2$ as ETL materials could lead to high-performance PSCs.

Fig. 4

Flow chart of the two-step ML in this work, which includes two steps, first-step ML (left side) and second-step ML (right side). Reproduced with permission from Ref. 80 © 2021 Royal Society of Chemistry.

Subsequently, to further improve the PCE, they focused on the impact of doping ETLs. A second dataset was created, comprising 90 device data points for PSCs based on doped ${\mathrm{SnO}}_2$ and 96 for PSCs based on doped ${\mathrm{TiO}}_2$. The physical and chemical properties and concentrations of the doping elements were used as model input features. They calculated the efficiency improvement ratio (EIR), which is the PCE of PSCs with doped ETL divided by the PCE of corresponding PSCs with undoped ETL, and used EIR as the model output. Using the RF model and genetic algorithms (GAs) for data training, they predicted doped ETLs that might lead to high EIR. Finally, based on predictions, they achieved PCEs of up to 30.47% and 28.54% for Cs-doped ${\mathrm{TiO}}_2$ ETL in CsFAMA-based PSCs and S-doped ${\mathrm{SnO}}_2$ ETL in FAMA-based PSCs, respectively.

3.1.4.

Optimizing lead-free perovskite

In order to develop environmentally friendly lead-free PSCs, ML is utilized to optimize the device architecture and components of Sn-based PSCs. Bak et al. initially collected real experimental data from 122 different photovoltaic devices with the perovskite structure ${\mathrm{ASnI}}_3$ from 49 published journal articles. The data include information on perovskite chemical structure, metal electrodes, transparent electrodes, HTL, and ETL as input features, while the corresponding short-circuit current density ($J_{\mathrm{SC}}$), open-circuit voltage ($V_{\mathrm{OC}}$), and fill factor (FF) values are considered as output features. A deep neural network (DNN) algorithm was then developed, and data augmentation was performed using K-fold cross-validation to randomly automate hyperparameter optimization of the DNN model to obtain the model’s pre-trained weights. The performance of the model was assessed by comparing the experimental data’s PCE values with the DNN’s predicted values for each layer (i.e., perovskite layer, metal electrode, transparent electrode, HTL, and ETL). Finally, by random combination of a large number of input features, predictions were made through the optimized DNN model (Fig. 5). ML recommended a new Sn-based PSC device structure $\mathrm{FTO}/\mathrm{PEDOT}:\mathrm{PSS}/{\mathrm{EDA}}_{0.01}{\mathrm{PEA}}_{0.07}{\mathrm{Cs}}_{0.03}{\mathrm{FA}}_{0.51}{\mathrm{MA}}_{0.38}{\mathrm{SnI}}_3/\text{indene}$-C60 bisadduct (ICBA)/2,9-dimethyl-4,7-diphenyl-1,10-phenanthrolin (BCP)/Ag. Obviously, the composition of the perovskite is still extremely complex. Furthermore, experimental verification has confirmed that the PCE of the ML-optimized Sn-based perovskite photovoltaics reached 5.57%, which is a significant improvement over the traditional testing results (PCE of 1.72%).⁸⁴

Fig. 5

Schematics of Sn-PSC design ML recommendation. (a) Parameter setting of each feature for the recommendation process. (b) Schematic of recommended Sn PSCs. (c) Labeled number and corresponding materials used as other layers in (a). Reproduced with permission from Ref. 84 (CC-BY).

3.1.5.

Vapor-deposited perovskite solar cells

Although the performance of solution-processed perovskite photovoltaics has rapidly improved in laboratory settings, scaling up to large-scale production remains a significant challenge. Vapor-deposited PSCs, which need to be compatible with the existing large-scale electronic industry, have substantially lagged behind. Wang et al. extracted information on 220 devices from 150 published papers on thermal evaporation PSCs between 2015 and 2022. Initially, the data were cleansed, reducing the sample size to 180 groups, and missing values in the dataset were filled with the mode of each column. The dataset was then expanded using Gaussian noise, tripling the sample size. Twenty features were chosen as input variables, encompassing aspects such as the material composition of the perovskite layer and device fabrication methods, with the device’s $J_{\mathrm{SC}}$, $V_{\mathrm{OC}}$, FF, and PCE as output variables. Training was conducted using 10 different ML models, with the RF model scoring highest on the test set. To enhance the generalization ability of the ML model, it was combined with a GA, increasing the prediction accuracy of the GA-enhanced RF model on the test set from 60.2% to 67.4%. Subsequently, the importance of features influencing PCE was analyzed through the mean Gini index of feature contributions in each decision tree of the RF model. The study found that the five features contributing most to the PCE were the proportions of MA, FA, Br, annealing temperature, and pressure (Fig. 6). Finally, the optimal device structure proposed was indium tin oxide $(\mathrm{ITO})/{\mathrm{SnO}}_2/\mathrm{BCP}/{\mathrm{FA}}_{0.25}{\mathrm{MA}}_{0.25}{\mathrm{Cs}}_{0.5}\mathrm{Pb}({\mathrm{I}}_{0.75}{\mathrm{Br}}_{0.25}{)}_3/\mathrm{MeO}\text{-}2\mathrm{PACz}/\mathrm{Ag}$, achieving the highest PCE of 26.1% under conditions of 90°C to 105°C and an evaporation pressure of $3\times {10}^{-5}$ to $4\times {10}^{-5}\mathrm{mbar}$. This work not only identified the optimal conditions for vapor-deposited perovskite but also, more crucially, delineated the direction for vapor-deposited perovskite.⁸⁵

Fig. 6

ML flowchart for studying the variables that influence the performance of vapor-deposited PSCs, consisting of four main parts: building the dataset, feature engineering, SHAP interpretation model, and model prediction. Reproduced with permission from Ref. 85 © 2023 Royal Society of Chemistry.

3.1.6.

Machine vision for large-area slot-die coated films

In the commercialization of perovskite photovoltaics, the controllable fabrication of large-area homogeneous perovskite films represents a critical technological challenge. Traditionally, the assessment of uniformity in large-area films has relied primarily on visual inspection, which is imprecise. Taherimakhsousi et al. developed a rapid, reliable, and non-destructive machine vision approach to quantify the uniformity of large-area perovskite films, thereby aiding process optimization.⁸⁶ Initially, white light photography was employed to sample slot-die coated perovskite films. Subsequently, a convolutional neural network based on the VGG16 architecture successfully segmented the original images into fully covered,⁸⁷ partially covered, and uncovered regions without the need for manual intervention. Optimized for pixel resolution, this approach spatially quantified multiple attributes of perovskite films (such as substrate coverage, film thickness, and defect density) from $25\mathrm{c}{\mathrm{m}}^2$ samples and established correlations between slot-die coating process parameters (wet film thickness and gas knife speed) and film properties (Fig. 7). This strategy facilitated multi-parameter, multi-objective optimization of the gas-knife assisted slot-die coating process, enabling the identification of optimal process conditions that simultaneously maximize coating throughput, film quality, and predicted device current density while reducing the characterization effort required for each process condition. This work demonstrates the significant potential of machine vision in optimizing large-area photovoltaic films.

Fig. 7

Machine vision workflow for quantifying large-area perovskite film morphology from optical images using the PerovskiteVision tool. Reproduced with permission from Ref. 86 (CC-BY).

3.2.1.

Predicting additive molecules

The quality of the perovskite active layer is crucial for the performance of perovskite LEDs. Currently, enhancing the quality of perovskite films involves additive engineering, which includes introducing organic molecules into the perovskite precursor solution. These molecules, while not integrating into the perovskite lattice, improve film quality by tuning the crystallization pathway. Yet, not every organic molecule is a suitable additive, and unsuitable choices can impair LED performance. Traditional additive selection based on trial-and-error is inefficient and challenged by the small size of perovskite LED additive databases for conventional ML. To address this, the enhanced molecular information model (EMIM) was developed, incorporating both qualitative molecular fingerprints and quantitative descriptors to enhance molecular information and improve predictive accuracy (Fig. 8).³⁹ It is noteworthy that in this work, the additive molecular database used was validated through experimental lab device tests. By utilizing EQE as the output feature and additive molecular descriptors as inputs, along with a 10.3% EQE threshold for classification, the EMIM model attained an impressive 98% prediction accuracy, significantly outperforming traditional ML models. In addition, among the 12 chosen additives, only one showed inconsistent results between EMIM predictions and device fabrication outcomes, demonstrating EMIM’s reliability. One of the additives even elevated the EQE of perovskite LEDs to 22.7%, making it one of the most efficient near-infrared LEDs. This work is pioneering in using ML for analyzing perovskite LEDs, especially for small data additives.

Fig. 8

Traditional method versus ML-assisted method. The traditional method is to verify the effectiveness of the additive for PeLEDs, while the ML-assisted method is to fabricate the device with the “good” additive predicted by ML. The additive is added into the perovskite precursor solution to form the perovskite film. Reproduced with permission from Ref. 39 © 2022 Wiley-VCH.

3.2.2.

Half-lifetime prediction

Stability is a key objective in the development of perovskite LEDs. Stability measurements, unlike EQE measurements, are time-consuming, necessitating hundreds of hours for the most stable perovskite LEDs. Therefore, developing strategies to predict the half-life of perovskite LEDs is crucial. Recently, we found that traditional exponential fitting methods, such as monoexponential, biexponential, and stretched exponential functions, are inadequate for predicting half-life due to complex degradation mechanisms involving ion migration, Joule heating, and interfacial effects.⁴⁰ To address this, an ensemble learning model was developed to predict the $T_{50}$ lifetime (time to 50% of initial performance) based on shorter tests $T_{80}$ (time to 80% of initial performance) (Fig. 9). A database of 210 near-infrared and red perovskite LEDs was initially compiled from experimental data. To augment this dataset, a DNN combined with the fast gradient sign method was employed, effectively doubling the dataset to 420 samples. Comparative analysis of five different ML models, such as least absolute shrinkage and selection operator, SVR, gradient boosting regressor, Gaussian process regressor, and elastic net (EN), showed improved accuracy over the original data. An ensemble learning model (5EML) based on these five algorithms was then used for better fitting the degradation curves of perovskite LEDs, achieving a high score of 0.995. This model also applies to quantum dot LEDs, demonstrating its universality.

Fig. 9

Illustration of lifetime prediction by 5EML. Reproduced with permission from Ref. 40 © 2024 Wiley-VCH.