Next Article in Journal
Laminar Biomaterial Composite of PVA Cryogel with Amnion as Potential Wound Dressing
Next Article in Special Issue
Recent Progress in Printed Photonic Devices: A Brief Review of Materials, Devices, and Applications
Previous Article in Journal
Chitosan-Based Nano Systems for Natural Antioxidants in Breast Cancer Therapy
Previous Article in Special Issue
Proposal and Numerical Analysis of Organic/Sb2Se3 All-Thin-Film Tandem Solar Cell
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Probing the Effect of Photovoltaic Material on Voc in Ternary Polymer Solar Cells with Non-Fullerene Acceptors by Machine Learning

1
College of Railway Transportation, Hunan University of Technology, Zhuzhou 412008, China
2
College of Electrical and Information Engineering, Hunan University of Technology, Zhuzhou 412008, China
3
Qinghai Provincial Key Laboratory of Nanomaterials and Nanotechnology, Qinghai Minzu University, Qinghai 810007, China
4
Shandong Provinical Key Laboratory of Optical Communication Science and Technology, School of Physical Science and Information Technology, Liaocheng University, Liaocheng 252059, China
*
Authors to whom correspondence should be addressed.
Submission received: 24 May 2023 / Revised: 28 June 2023 / Accepted: 29 June 2023 / Published: 5 July 2023
(This article belongs to the Special Issue Polymers for Electronics and Energy Devices)

Abstract

:
The power conversion efficiency (PCE) of ternary polymer solar cells (PSCs) with non-fullerene has a phenomenal increase in recent years. However, improving the open circuit voltage (Voc) of ternary PSCs with non-fullerene still remains a challenge. Therefore, in this work, machine learning (ML) algorithms are employed, including eXtreme gradient boosting, K-nearest neighbor and random forest, to quantitatively analyze the impact mechanism of Voc in ternary PSCs with the double acceptors from the two aspects of photovoltaic materials. In one aspect of photovoltaic materials, the doping concentration has the greatest impact on Voc in ternary PSCs. Furthermore, the addition of the third component affects the energy offset between the donor and acceptor for increasing Voc in ternary PSCs. More importantly, to obtain the maximum Voc in ternary PSCs with the double acceptors, the HOMO and LUMO energy levels of the third component should be around (−5.7 ± 0.1) eV and (−3.6 ± 0.1) eV, respectively. In the other aspect of molecular descriptors and molecular fingerprints in the third component of ternary PSCs with the double acceptors, the hydrogen bond strength and aromatic ring structure of the third component have high impact on the Voc of ternary PSCs. In partial dependence plot, it is clear that when the number of methyl groups is four and the number of carbonyl groups is two in the third component of acceptor, the Voc of ternary PSCs with the double acceptors can be maximized. All of these findings provide valuable insights into the development of materials with high Voc in ternary PSCs for saving time and cost.

1. Introduction

In recent years, polymer solar cells (PSCs) have become a hot spot in the photovoltaic field for integration with buildings due to their superior flexibility, light weight and high optical stability [1,2,3,4]. At present, the power conversion efficiency (PCE) of PSCs has exceeded 18% [5], which is mainly due to synthesizing new non-fullerene acceptors (NFAs), using new liquid/solid additives, developing layer-by-layer (LbL) fabrication process and applying the ternary strategy [6,7,8,9,10]. However, the Voc in the ternary PSCs is still lower than that of perovskite solar cells [11]. Some methods have been delivered, including effectively adjusting the NFA energy levels, to form the good energy level offset with the donor material for improving the Voc in devices [12,13]. Yong Cui et al. incorporated F-BTA3 as the third component in the solar cell with PBQx-TF:eC9-2Cl to form the cascaded energy level, which had a significant increase of Voc from 0.86 V to 0.879 V, meanwhile achieving a maximum PCE of 19.0% in ternary PSC [14]. Junyan Yang et al. introduced the small molecule donor of BTID-2F into PM6:Y6 to obtain lower energy loss, resulting in a Voc increase from 0.84 V of binary PSCs to 0.85 V of ternary PSCs, and PCE increased from 16.62% to 17.98%, which is attributed to the suppressed recombination and morphological optimization [15]. Guanshui Xie’s team also reported that adding the polymer donor of P1, with a deeper highest occupied molecular orbital (HOMO) energy level into the PM6:Y6 for reducing the HOMO level offset between Y6 and P1, which helped to decrease the energy loss and thus increase the Voc in PSCs [16]. These examples illustrate that adding the third component can effectively improve the Voc and PCE of ternary PSCs. However, the current research is basically based on the qualitative relationship studies in specific material system, and the quantitative analysis of the influence of the electronic properties and chemical structure of materials on PSCs’ Voc remains to be further studied.
Currently, density functional theory calculation, molecular dynamics simulation, multi-scale simulation model of semiconductor devices and empirical formula derivation have been applied for investigating the quantitative influence on PSC performance [17,18,19]. These research methods have important guidance for material screening and device optimization in solar cells. However, they often consume a lot of time and financial cost. Machine learning (ML) allows researchers to excavate the hidden physical laws behind big data and draw reliable conclusions from the perspective of algorithms for more effectively accelerating material design and performance optimization in PSCs to save time and money [20,21,22]. Currently, ML has been widely applied in the field of PSCs [23,24,25,26,27]. Kakaraparthi Kranthiraja et al. developed a series of novel π-conjugated polymer donor materials for NFAs based on the random forest (RF) model [23]. Sijing Zhong et al. reviewed the key challenges of ML in the design and improvement of PCE in photovoltaic materials in recent years and gave the guiding suggestions for reaching the high performance in devices [24]. Jinliang Wang et al. used ML to screen out 15 acceptors to match PTB7-Th for gaining more than 13% PCE [17]. In addition, our group applied ML to predict the champion PCE of ternary PSCs and screen the optimal doping ratio of the third component [26]. Furthermore, we also elucidated the matching energy levels of binary PSCs materials to enhance the Voc through ML [27]. Exploring organic materials and understanding hidden chemical information via ML have become a new research model. So far, the Voc in many classical binary PSCs still highly depends on the band gap between the HOMO energy level of the donor and the lowest unoccupied molecular orbital (LUMO) energy level of the acceptor [28]. However, the addition of the third component makes the mapping relationship between the energy level of materials and the Voc more complicated, resulting in the unclear influence mechanism of the Voc in ternary PSCs. In addition, the selection of the third component for high-efficiency ternary PSCs with non-fullerene often requires trial-and-error experiments. Therefore, there is still plenty of research space for utilizing ML to reveal the intrinsic relationship among the physical and chemical properties of materials and the Voc.
Herein, we analyze the important impact factors of the Voc in ternary PSCs with double-acceptors and explore the hidden correlation between the characteristics of non-fullerene molecules and Voc in ternary PSCs with double-acceptors via frontier molecular orbitals (FMOs), molecular descriptors (MDs) and molecular fingerprints (MFs). To our best knowledge, there is no research report to investigate the effect of MDs and MFs in non-fullerene materials on Voc in ternary PSCs with double-acceptors, and the overall process of ML is shown in Figure 1, which includes dataset collection, model building, performance evaluation and visual interpretation. The detail of the whole process is displayed in Materials and Methods. Moreover, Table S1 defines all of abbreviations, symbols, compounds, etc., in this work. We employ three different ML algorithms including K-nearest neighbors (KNN), RF and eXtreme gradient boosting (XGBoost) to investigate the complex relationship between FMOs of materials and Voc, and we also use the preferred XGBoost to study the influence of the physical and chemical properties of the third component on Voc. Additionally, the shapely additive explanations (SHAP) are used to analyze the influence of different input features in the XGBoost model and quantify the contribution of each input feature for Voc. In addition, combined with the division of the chemical structure in the third component material by MDs and MFs, the important influence of methyl and carbonyl functional groups on Voc is identified. It is worth mentioning that the relative amounts of methyl and carbonyl groups are quantified through partial dependence plots (PDP). Therefore, our results provide a guideline for the preparation of ternary PSCs with high Voc: (1) the optimal range of the third component energy level, (2) the visual relationship between molecular fragments of NFA materials and ternary PSCs’ Voc and (3) the marked key molecular functional groups for enhancing Voc.

2. Materials and Methods

2.1. Dataset Collection

All data are from 109 publications of non-fullerene ternary PSCs in the last five years, which are recorded as Data1 in the Supporting Information. Data1 contains seven input features (HOMO and LUMO of the donor (HOMO(D) and LUMO(D)), HOMO and LUMO of the acceptors (HOMO(A) and LUMO(A)), HOMO and LUMO of the third component (HOMO(T) and LUMO(T)), the concentration of the third component (T(%)) and one output feature (Voc)) with a total of 727 sets of data points. In part 3.1 of this paper, when studying the influence FMOs of donor and acceptor on Voc, the FMOs and the T(%) are used as input features, and Voc is the output feature. In part 3.2 of this work, when researching the effects of molecular composition and molecular properties on Voc, the third component of Data1 is trans-coded into two new sub-datasets via the ChemDes platform, which are the two-dimensional MDs and MFs, respectively. In addition, C_Voc is the Voc in the binary PSCs without the third component as the extra input feature, and the output feature of Voc is obtained at the value corresponding to the maximum PCE in the ternary PSCs.

2.2. Model Building

The ML models are built based on the Python language environment. The prediction model is established by K-nearest neighbors (KNN), random forest (RF) and eXtreme gradient boosting (XGBoost) algorithms. Among them, the KNN algorithm is a kind of non-parametric statistical lazy algorithm. For predicting the target value, the average of the K nearest values is assigned. In other words, similar inputs have similar outputs. The RF algorithm is a very representative bagging ensemble algorithm. All of the base evaluators are the decision trees. The XGBoost algorithm is built by the parallel construction of regression trees through multi-threading, which can control the model during the entire training process through a series of hyperparameters to greatly improve the speed and accuracy of model training. In addition, 20% of the total data in Table S5 is randomly selected as the testing set, while the remaining 80% is used as the training set (the training set and the testing set are mutually exclusive), and then the model is trained by using three type of ML algorithms. Random sampling ensures a good balance between the training and testing sets for preventing overfitting of the models. Then, the performance of the model is evaluated based on the predicted results. The model with the best performance results is selected for the predictive interpretation and the subsequent studies.

2.3. Performance Evaluation

As shown in Equations (1)–(4), the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and Pearson correlation coefficient (r) are used to evaluate the model with regression problems:
R M S E = i = 1 n ( y i y i ) 2 n
M A E = 1 n i = 1 n | ( y i y i ^ ) |
M A P E = 100 % n i = 1 n | y ^ i y i y i |
r = Σ ( x m x ) ( y m y ) Σ ( x m x ) 2 ( y m y ) 2
where n is the total numbers of data. yi and yi′ represent the measured Voc and the predicted Voc, respectively. ŷ and ŷ′ are the averages of the measured Voc and the predicted Voc, respectively. mx and my stand for the means of the x variable and the y variable, respectively. The higher r value indicates that the predicted Voc is in good agreement with the measured Voc. When the predicted Voc is in perfect agreement with the measured Voc, the RMSE is close to 0, and the r is close to 1, which reveals the model is better.

2.4. Visual Interpretation

SHAP originates from cooperative game theory and is gradually applied to explain various ML models, which makes the ML model not to be a “black-box model”. SHAP is an accumulated interpretive tool. The biggest advantage of SHAP is that it reflects the impact of each input feature for the output feature and also shows the negative or positive contribution of each input feature. In this work, we made three types of explanations for SHAP. (1) Global interpretation: The importance of all input features is sorted, and the contribution of each input feature for the output target (the predicted Voc) is visualized. (2) Local explanations: explaining the impact of features on the predicted Voc under a single feature or two feature interactions. (3) Sample interpretation: the explanatory analysis of the main input features for the individual sample predictive value in the output target (the predicted Voc).

3. Result and Discussions

3.1. The Influence FMOs of Donor and Acceptor on Voc

To analyze the correlation between FMOs of the donor and acceptor and the Voc in ternary PSCs, the Pearson correlation coefficient is calculated to measure the linear relationship between the input feature and output feature. As shown in Figure 2, it is not difficult to find that there is the highest correlation between Voc and the doping concentration T(%) of the third component (rT(%) = 0.282). Therefore, it is essential to focus on T(%) of the third component for increasing the Voc of the ternary PSCs. Tao Wang’s team reported that Voc gradually increased with the doping ratio concentration of IDMIC-4F into the PBDB-T-2F:BTP-4F binary PSCs. When the binary system PBDB-T-2F:BTP-4F has 0%, 5%, 10%, 15% and 20% of IDMIC-4F, respectively, the Voc of the device is 0.855 V, 0.858 V, 0.864 V, 0.867 V, 0.876 V and 0.890 V, respectively. The maximum PCE of 16.6% can be achieved with the Voc of 0.864 V, short circuit current (Jsc) density of 25.8 mA cm−2 and the Fill factor (FF) of 74.4% [29]. Increasing a third component can reduce non-radiative voltage loss in ternary PSCs, resulting in the higher Voc in devices [30,31]. More importantly, introducing a third component can significantly optimize the morphology of the active layer in ternary PSCs, which simultaneously promotes Jsc, Voc and FF [32,33]. Moreover, no matter whether it is the donor or the acceptor, the HOMO energy level shows the negative correlation with Voc, while the LUMO energy level is positively correlated with Voc in Figure 2. They reveal that the HOMO energy level of the donor (HOMO(D)) should be appropriately reduced or the LUMO energy level of the acceptor (HOMO(A)) should be increased to obtain a high Voc of ternary PSCs. On this basis, the introduction of the third component changes the arrangement of the energy level in the donor and acceptor, which can promote the exciton dissociation, improve the charge mobility and inhibit the recombination of the active layer to gain high Voc in ternary PSCs [34,35]. In addition, the correlation between Voc and the HOMO and LUMO of the third component (HOMO(T) and LUMO(T)) indicates that reducing HOMO(T) and LUMO(T) is beneficial to increase Voc. The reason for this may be that the cascaded energy level alignment of ternary PSCs accelerates the generated exciton dissociation process by multiple interfaces (Donor—Third component, Acceptor—Third component, Donor—Acceptor) to achieve a high Voc [36].
To assess the accuracy of the model, the XGBoost, KNN and RF algorithms based on the FMOs dataset are used to build the Voc prediction model. As seen in Figure 3, the three ML prediction models display that there is no significant outlier in the prediction results, indicating the potential of the models for predicting the Voc of ternary PSCs based on FMOs. Additionally, XGBoost and RF exhibit better accuracy than KNN in Figure 3a–c. In addition, Figure 3d shows the deviation from the predicted Voc to the measured Voc in three ML prediction models. It is clear that the primary counts of the deviation concentrate between 0 and 0.04 V, and deviations above 0.1 V are rare. They reveal that the predicted value of Voc is highly consistent with the measured value. Notably, the XGBoost, KNN and RF prediction models do not show overfitting, and the detailed performance evaluation is shown in Figure S1 of the Supporting Information. The XGBoost model has the best performance on the test set with RMSE (0.031), r (0.884), and MAE (0.022) at their best values, with an exception of MAPE(0.025), which is higher in XGBoost than in kNN(0.021), indicating that the XGBoost model has a great advantage in predicting Voc. Therefore, the XGBoost model is selected for further research.
In order to probe the underlying relationship between Voc and the FMOs of each material in the active layer of ternary PSCs, the SHAP method is employed based on the XGBoost algorithm to analyze the input feature importance and explain the contribution of each input feature for Voc. Figure 4a displays the importance score of all input features. It is easy to find that T(%) of the third component is the top factor to affect the Voc of ternary PSCs with non-fullerene, and the second factor is HOMO(A). In Figure 4b, the density scatter plot based on the SHAP interpreter reveals the contribution of each input feature on the Voc. Red stands for high eigenvalues, while low eigenvalues are represented by blue color. When the T(%) of the third component increases, the SHAP value increases in a positive direction, while HOMO(A) is the opposite, which is consistent with the previous conclusion in Figure 2. A deeper-lying HOMO level of the main acceptor favors lower energy losses to obtain a higher Voc in ternary PSCs [37,38,39]. Therefore, the main acceptor materials with low HOMO levels should be selected to improve the Voc of the ternary PSCs with two acceptors.
To investigate the impact of a single input feature on the Voc of ternary PSCs, the SHAP dependency relationship is utilized to analyze the contribution of a single input feature to the prediction results. It should be noted that SHAP emphasizes single sample interpretation, and the contribution of each input feature is not directly related to the amount of data. Figure 4 illustrates the results of the analysis on the testing set. The red dotted box denotes the area with a positive SHAP value, which stands for promoting Voc. Additionally, each point is colored to highlight the interaction with input feature. Figure 5a,b reveal that higher-lying LUMO(T) has more of a contribution to Voc, while Figure 5c,d show that the Voc has an upward trend with the lower-lying HOMO(T). However, all changed values of LUMO and HOMO in the third component must be appropriate. As shown in Figure 5a,b, most of the SHAP values are in the positive range, resulting in the increased Voc with the high-level LUMO(T), but there are not too many changes of Voc with the gradually increased LUMO(T). This underscores the significance of energy level matching for achieving an effective improvement in Voc of ternary PSCs [40]. Similarly, as shown in Figure 5c,d, the same trend happens for the effect of the deeper HOMO(T) on Voc of ternary PSCs. More importantly, the color rendering from Figure 5 shows that the co-effect of the HOMO energy level of the third component (HOMO(T)) with HOMO(D) and HOMO(A) on Voc being greater than that of LUMO(T) with the LUMO energy level of the donor (LUMO(D)) and the LUMO energy level of the acceptor (LUMO(A)) on Voc. Therefore, Voc in ternary PSCs has a high relationship with the third component, because it changes the matched energy level of the donor and acceptor, leading to the reduction of the offset between the energy levels of donor and acceptor. Lower-lying HOMO(T) is conducive to the construction of the D:A1:T cascaded energy level structure, while higher-lying LUMO(T) reduces the band gap for increasing Voc in ternary PSCs [41]. Additionally, it is interesting that the main positive SHAP value of HOMO(T) is around (−5.7 ± 0.1) eV and the most positive SHAP value of LUMO(T) is around (−3.6 ± 0.1) eV. Kuibao Yu et al. selected Y6-1O as the third component, which has a similar molecular structure and energy level as the Y6 acceptor, and the HOMO/LUMO energy levels of Y6 and Y6-1O are −5.71/−4.10 eV and −5.71/−3.84 eV, respectively. Due to the high compatibility between the two acceptors, the third component Y6-1O is completely embedded in the main acceptor Y6 to form an alloying phase. It makes the energy transfer effect between them significantly enhanced. The Voc of ternary PSCs is increased with the additive of Y6-1O content, which is attributed to the fact that the optimized ternary PSCs enhance charge mobility and inhibit charge recombination [42]. The given energy level range of the third component (HOMO(T)) is around (−5.7 ± 0.1) eV. LUMO(T) is about (−3.6 ± 0.1) eV and can provide guidance to screen, design and synthesize suitable non-fullerene materials for ternary PSCs.
To explain the predicted Voc value with the synergistic effect of each input feature, the waterfall diagram in Figure 6 with the example of ternary PSCs based on PBDB-T-2F:BTP-4F:IDMIC-4F is applied to identify the quantitative contribution of each input feature to Voc [29]. The prediction results of the model are driven from the base value to the final model output according to the contribution of each input feature. The input features to increase the predicted value are shown in red, indicating a positive effect, while the input features to reduce the prediction are shown in blue, meaning a negative effect. It is not difficult to find that the input features of HOMO(A) and LUMO(T) reduce the Voc value, and the rest of the input features promote the improvement of Voc with the different degrees. Remarkably, T(%) is the most important factor for improving Voc. The SHAP values of the main acceptor’s HOMO and the third component’s LUMO are (−4.0 × 10−4) eV and (−2.6 × 10−3) eV, respectively. It may be because HOMO(T) (−5.46 eV) is higher than (−5.7 ± 0.1) eV so that it did not increase Voc synergistically with HOMO(A), while LUMO(T) is lower (−3.83 eV), which is consistent with the previous conclusion in Figure 5 [43]. The predicted results also partly explain the synergistically effect of each input feature on the Voc of ternary PSCs.

3.2. The Effect of MDs and MFs on Voc

Since the energy level of third component is very important for increasing the Voc in ternary PSCs and the energy level of material is highly dependent on the chemical structure of material, the electronic topological state (E-state) MDs and Klekota–Roth MFs of the third component materials are used to analyze the influence of chemical information of the third component materials on Voc. Firstly, the method of feature dimensionality reduction is used to select input features for avoiding the overfitting of ML, and the selected features are displayed in Figure 7a,b. The control the open circuit voltage (C_Voc) is the Voc of PSCs without the third component, and it has the highest correlation with Voc (rC_Voc = 0.847), which has a positive effect on the Voc improvement of the ternary PSCs. It reveals that introducing a third component to binary devices with high Voc is more likely to obtain a higher Voc in ternary PSCs. The selected 10 features of electrical topological states are described in Table S3. Moreover, the correlations between maxHB, SBint9 and Voc are 0.247 and 0.213, respectively, as depicted in Figure 7a, which suggests that the third component should have strong hydrogen bonding to enhance the Voc of ternary PSCs. Hydrogen bond strength can alter the morphology of the active layer and inhibit the bimolecular recombination for greatly improving Voc [44]. Furthermore, the correlation between nsCH3 and Voc is 0.191, indicating that a certain number of methyl groups plays a critical role in enhancing Voc. Notably, hydrogen atoms in the methyl group can form hydrogen bonds. However, their low electronegativity renders the hydrogen bonds strength in the methyl group to be relatively weak compared with the hydrogen bond strength with other atoms [45]. Therefore, comprehensively considering the interaction of hydrogen bond strength and the number of methyl groups of the third component is crucial for enhancing Voc in ternary PSCs.
In addition, eight key input features of the Klekota–Roth MFs with the strong effect on Voc are chosen. As seen in Figure 7b, among the selected fragments, KR1193 containing a carbonyl group is identified as the most influential feature on Voc. KR1193, KR4501, KR4156 and KR3349 in Klekota–Roth MFs contain the carbonyl structures, which are consistent with carbonyl characteristics from the E-state MDs. Their descriptions and structures are shown in Figure 8. It further verifies that the design of carbonyls in the third component molecule is very important for Voc. KR4501, KR4287, KR3568 and KR415 represent the aromatic ring connected with other different atoms, respectively. These fragments are used to synthesize new molecules, which have stronger intermolecular π-π packing and better backbone coplanarity, resulting in high Voc [46]. Moreover, these aromatic ring structures also promote electron transfer due to their large polarizability and formation of widely distributed molecular orbitals [47]. All in all, the design of NFA in the third component molecules with the aromatic ring fragments may be an effective strategy for improving Voc in ternary PSCs with dual acceptors.
To evaluate model performance and quantify the effect of functional groups on Voc in the third component to determine the appropriate number of important functional groups in the molecule, we used the XGBoost algorithm to predict the Voc of the ternary PSCs, as shown in Figure S2. The performance evaluation of the XGBoost algorithm is shown in Table S4 ( RMSE = 0.031, MAE = 0.019, MAPE = 0.023 and r = 0.832 in the test set of E-state MDs, and RMSE = 0.032, MAE = 0.019, MAPE = 0.022 and r = 0.822 in the test set of Klekota–Roth MFs). All results illustrate the feasibility of XGBoost in predicting Voc. Moreover, the partial dependence diagrams is utilized to investigate the relationship between the key molecular fragments in the third component and Voc, as seen in Figure 9a,b. The results indicate that increasing the number of methyl and carbonyl can enhance the Voc of ternary PSCs. Specifically, with the increase of the number of methyl, the quantitative value of contribution in Voc increases. When the number of methyl is greater than four, the Voc is no longer changed. Jianhui Hou’s team designed BTP-M by introducing a weakly electron-donating methyl group to replace the electron-withdrawing F atom at the terminal group into Y6. BTP-M exhibited a higher LUMO than Y6 and reduced the energy loss in ternary PSCs with PM6:Y6:BTP-M, leading to a higher Voc and larger Jsc in devices [36]. Moreover, methyl groups in different positions on the side chains can also cause differences in molecular properties [48]. The number and position of methyl groups are reasonably adjusted to synthesizing the NFA of the third component molecules with high performance and optimizing charge transport in the active layer. More importantly when the number of carbonyl groups is bigger than 2, Voc stays the same. Additionally, carbonyl groups can form non-covalent bonds with other groups, contributing to a better balance between Voc and Jsc in ternary PSCs. For example, Xinrui Li et al. found that in the ternary PSCs with PTB7-Th:SR197:ITIC, the carbonyl group of ITIC reacts with the N-H group of SR197 to form N−H…O non-covalent interactions, resulting in an increase in Jsc and FF while Voc is almost unchanged, with PCE ranging from 7.92% to 10.29% [49]. Thus, when designing the NFA of the third component molecules to improve the Voc of ternary PSCs, the number of carbonyl groups should be taken into consideration. Moreover, the experimental devices’ Voc values from the publication that is not in datasets are used to verify the predicted Voc by ML in Table S2 in the Supporting Information. It is not difficult to find that the predicted Voc by ML just has a small relative error compared with the measured Voc in the literature which is not in datasets. It reveals that the prediction ML model has certain reliability.

4. Conclusions

In this work, we have comprehensively investigated the influence of materials on the Voc of ternary PSCs with dual acceptors by two ways: one is the FMOs of the donor and acceptor, and the other is MDs and MFs in the third component molecules of NFAs. Firstly, three different ML algorithms, XGBoost, KNN and RF, are employed to predict the Voc of ternary PSCs with the energy level of active layer material as input features. In three models, XGBoost has the highest accuracy for predicting the Voc of ternary PSCs, with RMSE being 0.031, MAE being 0.022, MAPE being 0.025 and r being 0.884 in the test set. Secondly, the T(%) value of the third component has the most significant impact on Voc. It is worth mentioning that the HOMO and LUMO energy levels of the third component should be at (−5.7 ± 0.1) eV and (−3.6 ± 0.1) eV for gaining the maximum Voc value in ternary PSCs with two acceptors. Finally, based on the partial dependence analysis, the hydrogen bond strength and aromatic ring in the third component can enhance Voc, and the optimal third component of NFA acceptor molecule should have four methyl groups and two carbonyl groups which should be identified. This work provides design guidance for developing NFA materials and optimizing the Voc value of ternary PSCs.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/polym15132954/s1, Figure S1: The evaluation performance (a) RMSE, (b) MAE, (c) MAPE and (d) r of the training and testing sets based on three ML algorithms and FMOs as input feature descriptors in Data1. Three ML algorithms are XGBoost, KNN and RF, respectively; Figure S2: The relationship between the predicted Voc and measured Voc values based on the XGBoost model and (a) the E-state MDs and (b) Klekota–Roth MFs as input feature descriptors, respectively; Table S1: The nomenclature and full name of abbreviations, symbols, compounds, etc; Table S2: Device characteristics of the ternary PSCs in literature. And the relative error (%) between the predicted values of Voc based on XGBoost model and measured Voc in literature, which is calculated as = [(Voc predicted - Voc measured)/ Voc measured]*%; Table S3: The specific meaning of 9 selected molecular 2D E-states; Table S4: The evaluation performance by XGBoost based on 2D E-state MDs and Klekota-Roth MFs; Table S5: Data set of Machine Learning. References [50,51,52,53,54] are cited in the supplementary materials.

Author Contributions

Conceptualization, D.H. and L.Z.; Investigation, X.Z., X.P. and R.Z.; Methodology and Formal Analysis, Z.L., K.W. and H.Z.; Validation, Z.L., K.W. and J.W.; Visualization, Z.L. and K.W.; Writing—Original Draft, Z.L. and D.H.; Writing—Review and Editing, J.W., J.L. and L.Z.; Funding Acquisition, D.H. and J.L. All authors have read and agreed to the published version of the manuscript.

Funding

This study has received support of the fundings, including the Natural Science Foundation of Hunan Province (Grant No. 2021JJ40168), the Education Department of Hunan Province (Grant Nos. 22B0580 and 21B0534), the Open Fund Project of Qinghai Provincial Laboratory of Nanomaterials and Nanotechnology (Grant No. 2022-QHMD-PI-NM-03) and the National innovation and entrepreneurship training project for college students (Grant No. 202211535013).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The data are available in the Supporting Information.

Acknowledgments

The authors thank Qian Peng from China University of Labor Relations for polishing our manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Liu, H.; Wu, J.; Fu, Y.; Wang, B.; Yang, Q.; Sharma, G.D.; Keshtov, M.L.; Xie, Z. One-step solution-processed low surface roughness silver nanowire composite transparent electrode for efficient flexible indium tin oxide-free polymer solar cells. Thin Solid Film. 2021, 718, 138486. [Google Scholar] [CrossRef]
  2. Hu, L.; Song, J.; Yin, X.; Su, Z.; Li, Z. Research progress on polymer solar cells based on PEDOT:PSS electrodes. Polymers 2020, 12, 145. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  3. Sorrentino, R.; Kozma, E.; Luzzati, S.; Po, R. Interlayers for non-fullerene based polymer solar cells: Distinctive features and challenges. Energy Environ. Sci. 2021, 14, 180–223. [Google Scholar] [CrossRef]
  4. Lin, Y.-C.; Lu, Y.-J.; Tsao, C.-S.; Saeki, A.; Li, J.-X.; Chen, C.-H.; Wang, H.-C.; Chen, H.-C.; Meng, D.; Wu, K.-H. Enhancing photovoltaic performance by tuning the domain sizes of a small-molecule acceptor by side-chain-engineered polymer donors. J. Mater. Chem. A 2019, 7, 3072–3082. [Google Scholar] [CrossRef]
  5. Ahmad, N.; Zhou, H.; Fan, P.; Liang, G. Recent progress in cathode interlayer materials for non-fullerene organic solar cells. EcoMat 2022, 4, e12156. [Google Scholar] [CrossRef]
  6. Zhang, Z.; Shan, T.; Zhang, Y.; Zhu, L.; Kong, L.; Liu, F.; Zhong, H. Isomerizing thieno [3, 4-b] thiophene-based near-infrared non-fullerene acceptors towards efficient organic solar cells. J. Mater. Chem. C 2020, 8, 4357–4364. [Google Scholar] [CrossRef]
  7. Wang, H.; Cao, J.; Yu, J.; Zhang, Z.; Geng, R.; Yang, L.; Tang, W. Molecular engineering of central fused-ring cores of non-fullerene acceptors for high-efficiency organic solar cells. J. Mater. Chem. A 2019, 7, 4313–4333. [Google Scholar] [CrossRef]
  8. Gokulnath, T.; Gayathri, R.D.; Park, H.-Y.; Kim, J.; Kim, H.; Kim, J.; Reddy, S.S.; Yoon, J.; Jin, S.-H. Highly efficient layer-by-layer deposition solar cells achieved with halogen-free solvents and molecular engineering of non-fullerene acceptors. Chem. Eng. J. 2022, 448, 137621. [Google Scholar] [CrossRef]
  9. Lin, Y.-C.; Chen, C.-H.; Li, R.-H.; Tsao, C.-S.; Saeki, A.; Wang, H.-C.; Chang, B.; Huang, L.-Y.; Yang, Y.; Wei, K.-H. Atom-varied side chains in conjugated polymers affect efficiencies of photovoltaic devices incorporating small molecules. ACS Appl. Polym. Mater 2019, 2, 636–646. [Google Scholar] [CrossRef]
  10. An, Q.; Zhang, F.; Zhang, J.; Tang, W.; Deng, Z.; Hu, B. Versatile ternary organic solar cells: A critical review. Energy Environ. Sci. 2016, 9, 281–322. [Google Scholar] [CrossRef]
  11. Wu, Y.; Ding, N.; Zhang, Y.; Liu, B.; Zhuang, X.; Liu, S.; Nie, Z.; Bai, X.; Dong, B.; Xu, L. Toward broad spectral response inverted perovskite solar cells: Insulating quantum-cutting perovskite nanophosphors and multifunctional ternary organic bulk-heterojunction. Adv. Energy Mater. 2022, 12, 2200005. [Google Scholar] [CrossRef]
  12. Avalos-Quiroz, Y.A.; Bardagot, O.; Kervella, Y.; Aumaître, C.; Cabau, L.; Rivaton, A.; Margeat, O.; Videlot-Ackermann, C.; Vongsaysy, U.; Ackermann, J. Non-fullerene acceptors with an extended π-conjugated core: Third components in ternary blends for high-efficiency, post-treatment-free organic solar cells. ChemSusChem 2021, 14, 3502–3510. [Google Scholar] [CrossRef] [PubMed]
  13. Liu, Y.; Chen, Y. Integrated perovskite/bulk-heterojunction organic solar cells. Adv. Mater. 2020, 32, 1805843. [Google Scholar] [CrossRef] [PubMed]
  14. Cui, Y.; Xu, Y.; Yao, H.; Bi, P.; Hong, L.; Zhang, J.; Zu, Y.; Zhang, T.; Qin, J.; Ren, J. Single-junction organic photovoltaic cell with 19% efficiency. Adv. Mater. 2021, 33, 2102420. [Google Scholar] [CrossRef] [PubMed]
  15. Yan, Y.; Zhang, Y.; Liu, Y.; Shi, Y.; Qiu, D.; Deng, D.; Zhang, J.; Wang, B.; Adil, M.A.; Amin, K. Simultaneously decreasing the bandgap and voc loss in efficient ternary organic solar cells. Adv. Energy Mater. 2022, 12, 2200129. [Google Scholar] [CrossRef]
  16. Xie, G.; Zhang, Z.; Su, Z.; Zhang, X.; Zhang, J. 16.5% efficiency ternary organic photovoltaics with two polymer donors by optimizing molecular arrangement and phase separation. Nano Energy 2020, 69, 104447. [Google Scholar] [CrossRef]
  17. Mahmood, A.; Irfan, A.; Wang, J.-L. Machine learning and molecular dynamics simulation-assisted evolutionary design and discovery pipeline to screen efficient small molecule acceptors for PTB7-Th-based organic solar cells with over 15% efficiency. J. Mater. Chem. A 2022, 10, 4170–4180. [Google Scholar] [CrossRef]
  18. Sethi, S.K.; Singh, M.; Manik, G. A multi-scale modeling and simulation study to investigate the effect of roughness of a surface on its self-cleaning performance. Mol. Syst. Des. Eng. 2020, 5, 1277–1289. [Google Scholar] [CrossRef]
  19. Scharber, M.C.; Mühlbacher, D.; Koppe, M.; Denk, P.; Waldauf, C.; Heeger, A.J.; Brabec, C.J. Design rules for donors in bulk-heterojunction solar cells—Towards 10% energy-conversion efficiency. Adv. Mater. 2006, 18, 789–794. [Google Scholar] [CrossRef]
  20. Lee, M.-H. A machine learning-based design rule for improved open-circuit voltage in ternary organic solar cells. Adv. Intell. Syst. 2020, 2, 1900108. [Google Scholar] [CrossRef]
  21. Liu, X.; Shao, Y.; Lu, T.; Chang, D.; Li, M.; Lu, W. Accelerating the discovery of high-performance donor/acceptor pairs in photovoltaic materials via machine learning and density functional theory. Mater. Des. 2022, 216, 110561. [Google Scholar] [CrossRef]
  22. Malhotra, P.; Biswas, S.; Chen, F.-C.; Sharma, G.D. Prediction of non-radiative voltage losses in organic solar cells using machine learning. SoEn 2021, 228, 175–186. [Google Scholar] [CrossRef]
  23. Kranthiraja, K.; Saeki, A. Machine learning-assisted polymer design for improving the performance of non-fullerene organic solar cells. ACS Appl. Mater. Interfaces 2022, 14, 28936–28944. [Google Scholar] [CrossRef] [PubMed]
  24. Zhong, S.; Yap, B.K.; Zhong, Z.; Ying, L. Review on Y6-based semiconductor materials and their future development via machine learning. Crystals 2022, 12, 168. [Google Scholar] [CrossRef]
  25. Kranthiraja, K.; Saeki, A. Experiment-oriented machine learning of polymer: Non-fullerene organic solar cells. Adv. Funct. Mater. 2021, 31, 2011168. [Google Scholar] [CrossRef]
  26. Guo, C.; Li, Z.; Wang, K.; Zhou, X.; Huang, D.; Liang, J.; Zhao, L. Accelerated exploration of efficient ternary solar cells with PTB7: PC 71 BM: SMPV1 using machine-learning methods. Phys. Chem. Chem. Phys 2022, 24, 22538–22545. [Google Scholar] [CrossRef]
  27. Wang, K.; Guo, C.R.; Li, Z.N.; Zhang, R.; Feng, Z.M.; Fang, G.K.; Huang, D.; Liang, J.J.; Zhao, L.; Li, Z.C. Machine learning assisted identification of the matched energy level of materials for high open circuit voltage in binary organic solar cells. Mol. Syst. Des. Eng. 2023, 8, 799–809. [Google Scholar] [CrossRef]
  28. Nie, Q.; Tang, A.; Guo, Q.; Zhou, E. Benzothiadiazole-based non-fullerene acceptors. Nano Energy 2021, 87, 106174. [Google Scholar] [CrossRef]
  29. Li, D.; Chen, X.; Cai, J.; Li, W.; Chen, M.; Mao, Y.; Du, B.; Smith, J.A.; Kilbride, R.C.; O’Kane, M.E. Non-fullerene acceptor fibrils enable efficient ternary organic solar cells with 16.6% efficiency. Sci. China Chem. 2020, 63, 1461–1468. [Google Scholar] [CrossRef]
  30. Suthar, R.; Dahiya, H.; Karak, S.; Sharma, G.D. Ternary organic solar cells: Recent insight on structure-processing-property-performance relationships. Energy Technol. 2023, 11, 2201176. [Google Scholar] [CrossRef]
  31. Yin, Y.; Zhan, L.; Liu, M.; Yang, C.; Guo, F.; Liu, Y.; Gao, S.; Zhao, L.; Chen, H.; Zhang, Y. Boosting photovoltaic performance of ternary organic solar cells by integrating a multi-functional guest acceptor. Nano Energy 2021, 90, 106538. [Google Scholar] [CrossRef]
  32. Li, F.; Chen, Y.; Fan, X.-H.; Gao, C.-Y.; Zhu, X.; Yang, L.-M. High performance achieved via core engineering and side-chain engineering in organic solar cells based on the penta-fused-ring acceptor. J. Mater. Chem. C 2022, 10, 7724–7730. [Google Scholar]
  33. Zhang, Y.; Li, G. Functional third components in nonfullerene acceptor-based ternary organic solar cells. Acc. Chem. Res. 2020, 1, 158–171. [Google Scholar] [CrossRef]
  34. Wang, H.; Zhang, Z.; Yu, J.; Liu, X.; Tang, W. High mobility acceptor as third component enabling high-performance large area and thick active layer ternary solar cells. Chem. Eng. J. 2021, 418, 129539. [Google Scholar] [CrossRef]
  35. Ma, Q.; Jia, Z.; Meng, L.; Zhang, J.; Zhang, H.; Huang, W.; Yuan, J.; Gao, F.; Wan, Y.; Zhang, Z.; et al. Promoting charge separation resulting in ternary organic solar cells efficiency over 17.5%. Nano Energy 2020, 78, 105272. [Google Scholar] [CrossRef]
  36. Hao, T.; Leng, S.; Yang, Y.; Zhong, W.; Zhang, M.; Zhu, L.; Song, J.; Xu, J.; Zhou, G.; Zou, Y.; et al. Capture the high-efficiency non-fullerene ternary organic solar cells formula by machine-learning-assisted energy-level alignment optimization. Patterns 2021, 2, 100333. [Google Scholar] [CrossRef]
  37. An, Q.; Wang, J.; Ma, X.; Gao, J.; Hu, Z.; Liu, B.; Sun, H.; Guo, X.; Zhang, X.; Zhang, F. Two compatible polymer donors contribute synergistically for ternary organic solar cells with 17.53% efficiency. Energy Environ. Sci. 2020, 13, 5039–5047. [Google Scholar] [CrossRef]
  38. Lu, H.; Liu, W.; Jin, H.; Huang, H.; Tang, Z.; Bo, Z. High-efficiency organic solar cells with reduced nonradiative voltage loss enabled by a highly emissive narrow bandgap fused ring acceptor. Adv. Funct. Mater. 2022, 32, 2107756. [Google Scholar] [CrossRef]
  39. Bi, Z.; Naveed, H.B.; Wu, H.; Zhang, C.; Zhou, X.; Wang, J.; Wang, M.; Wu, X.; Zhu, Q.; Zhou, K. Tuning acceptor composition in ternary organic photovoltaics–impact of domain purity on non-radiative voltage losses. Adv. Energy Mater. 2022, 12, 2103735. [Google Scholar]
  40. Yu, R.; Yao, H.; Hou, J. Recent progress in ternary organic solar cells based on nonfullerene acceptors. Adv. Energy Mater. 2018, 8, 1702814. [Google Scholar] [CrossRef]
  41. Yang, H.; Dong, Y.; Fan, H.; Wu, Y.; Cui, C.; Li, Y. A Large-bandgap guest material enabling improved efficiency and reduced energy loss for ternary polymer solar cells. Sol. RRL 2021, 5, 2100013. [Google Scholar]
  42. Yu, K.; Song, W.; Li, Y.; Chen, Z.; Ge, J.; Yang, D.; Zhang, J.; Xie, L.; Liu, C.; Ge, Z. Achieving 18.14% efficiency of ternary organic solar cells with alloyed nonfullerene acceptor. Small Struct. 2021, 2, 2100099. [Google Scholar] [CrossRef]
  43. Raheem, A.A.; Murugan, P.; Shanmugam, R.; Praveen, C. Azulene bridged π-distorted chromophores: The influence of structural symmetry on optoelectrochemical and photovoltaic parameters. ChemPlusChem 2021, 86, 1451–1460. [Google Scholar] [CrossRef] [PubMed]
  44. Li, X.; Zhou, L.; Lu, X.; Cao, L.; Du, X.; Lin, H.; Zheng, C.; Tao, S. Hydrogen bond induced high-performance quaternary organic solar cells with efficiency up to 17.48% and superior thermal stability. Mater. Chem. Front. 2021, 5, 3850–3858. [Google Scholar] [CrossRef]
  45. Dutta, J.; Sahu, A.K.; Bhadauria, A.S.; Biswal, H.S. Carbon-centered hydrogen bonds in proteins. J. Chem. Inf. Model. 2022, 62, 1998–2008. [Google Scholar] [CrossRef]
  46. Gokulnath, T.; Choi, J.; Park, H.-Y.; Sung, K.; Do, Y.; Park, H.; Kim, J.; Reddy, S.S.; Kim, J.; Song, M. A wide-bandgap π-conjugated polymer for high-performance ternary organic solar cells with an efficiency of 17.40%. Nano Energy 2021, 89, 106323. [Google Scholar] [CrossRef]
  47. Chen, Y.; Yan, C.; Dong, J.; Zhou, W.; Rosei, F.; Feng, Y.; Wang, L.N. Structure/property control in photocatalytic organic semiconductor nanocrystals. Adv. Funct. Mater. 2021, 31, 2104099. [Google Scholar] [CrossRef]
  48. Zhang, K.-N.; Bi, P.-Q.; Wen, Z.-C.; Niu, M.-S.; Chen, Z.-H.; Wang, T.; Feng, L.; Yang, J.-L.; Hao, X.-T. Unveiling the important role of non-fullerene acceptors crystallinity on optimizing nanomorphology and charge transfer in ternary organic solar cells. Org. Electron. 2018, 62, 643–652. [Google Scholar] [CrossRef]
  49. Li, X.; Du, X.; Lin, H.; Kong, X.; Li, L.; Zhou, L.; Zheng, C.; Tao, S. Ternary system with intermolecular hydrogen bond: Efficient strategy to high-performance nonfullerene organic solar cells. ACS Appl. Mater. Interfaces 2019, 11, 15598–15606. [Google Scholar] [CrossRef]
  50. Wang, L.; Zhu, Z.; Yang, S.; Fan, J.; Huang, S.; Yang, S.; Li, H.; Liu, H. Boosting the performance of all-polymer solar cells via incorporating a versatile small-molecule non-fullerene acceptor. Synth. Met. 2023, 293, 117292. [Google Scholar] [CrossRef]
  51. Liu, X.; Liu, Y.; Ni, Y.; Fu, P.; Wang, X.; Yang, Q.; Guo, X.; Li, C. Reducing non-radiative recombination energy loss via a fluorescence intensifier for efficient and stable ternary organic solar cells. Mater. Horizons 2021, 8, 2335–2342. [Google Scholar] [CrossRef]
  52. Lan, A.; Lv, Y.; Zhu, J.; Lu, H.; Do, H.; Chen, Z.-K.; Zhou, J.; Wang, H.; Chen, F.; Zhou, E. High-performance ternary organic solar cells through incorporation of a series of A2-A1-D-A1-A2 type nonfullerene acceptors with different terminal groups. ACS Energy Lett. 2022, 7, 2845–2855. [Google Scholar]
  53. Huang, T.; Zhang, Z.; Wang, D.; Zhang, Y.; Deng, Z.; Huang, Y.; Liao, Q.; Zhang, J. 18.7% Efficiency Ternary Organic Solar Cells Using Two Non-Fullerene Acceptors with Excellent Compatibility. ACS Appl. Energy Mater. 2023, 6, 3126–3134. [Google Scholar] [CrossRef]
  54. Chen, T.; Li, S.; Li, Y.; Chen, Z.; Wu, H.; Lin, Y.; Gao, Y.; Wang, M.; Ding, G.; Min, J.; et al. Compromising Charge Generation and Recombination of Organic Photovoltaics with Mixed Diluent Strategy for Certified 19.4% Efficiency. Adv. Mater. 2023, 35, e2300400. [Google Scholar] [CrossRef]
Figure 1. The machine learning process used in this work. (a) Dataset collection, (b) model building, (c) performance evaluation and (d) visual interpretation. Additionally, PDP stands for partial dependence plots, and SHAP represents the shapely additive explanations in the illustration.
Figure 1. The machine learning process used in this work. (a) Dataset collection, (b) model building, (c) performance evaluation and (d) visual interpretation. Additionally, PDP stands for partial dependence plots, and SHAP represents the shapely additive explanations in the illustration.
Polymers 15 02954 g001
Figure 2. Pearson correlation matrix between Voc and input features. (The value of the correlation coefficient r represents the direction and strength of the linear relationship between the input and output feature, which range from −1 to +1, and the corresponding color is from blue to red.)
Figure 2. Pearson correlation matrix between Voc and input features. (The value of the correlation coefficient r represents the direction and strength of the linear relationship between the input and output feature, which range from −1 to +1, and the corresponding color is from blue to red.)
Polymers 15 02954 g002
Figure 3. The relationship between the predicted Voc and measured Voc based on different regression algorithms (a) XGBoost, (b) KNN and (c) RF (black dots represent training data; red dots stand for testing data, and when data points fall on green dash lines, it means that the predicted and measured Voc are equal), (d) Histogram of frequency distribution of fitting error in the XGBoost, KNN and RF models.
Figure 3. The relationship between the predicted Voc and measured Voc based on different regression algorithms (a) XGBoost, (b) KNN and (c) RF (black dots represent training data; red dots stand for testing data, and when data points fall on green dash lines, it means that the predicted and measured Voc are equal), (d) Histogram of frequency distribution of fitting error in the XGBoost, KNN and RF models.
Polymers 15 02954 g003
Figure 4. (a) Ranking of input feature importance, (b) SHAP value summary plot for the FMOs of each material based on XGBoost algorithm.
Figure 4. (a) Ranking of input feature importance, (b) SHAP value summary plot for the FMOs of each material based on XGBoost algorithm.
Polymers 15 02954 g004
Figure 5. (a) The dependent relationship on LUMO(D) and its SHAP values with the changes of LUMO(T), (b) the dependent relationship on LUMO(A) and its SHAP values with the changes of LUMO(T), (c) the dependent relationship on HOMO(D) and its SHAP values with the changes of HOMO(T) and (d) the dependent relationship on HOMO(A) and its SHAP values with the changes of HOMO(T).
Figure 5. (a) The dependent relationship on LUMO(D) and its SHAP values with the changes of LUMO(T), (b) the dependent relationship on LUMO(A) and its SHAP values with the changes of LUMO(T), (c) the dependent relationship on HOMO(D) and its SHAP values with the changes of HOMO(T) and (d) the dependent relationship on HOMO(A) and its SHAP values with the changes of HOMO(T).
Polymers 15 02954 g005
Figure 6. The waterfall diagram of the predicted Voc for PBDB-T-2F:BTP-4F:IDMIC-4F.
Figure 6. The waterfall diagram of the predicted Voc for PBDB-T-2F:BTP-4F:IDMIC-4F.
Polymers 15 02954 g006
Figure 7. (a) Pearson correlation matrix of Voc and E-state MDs and (b) Klekota–Roth MFs.
Figure 7. (a) Pearson correlation matrix of Voc and E-state MDs and (b) Klekota–Roth MFs.
Polymers 15 02954 g007
Figure 8. The screened 10 Klekota–Roth MF features.
Figure 8. The screened 10 Klekota–Roth MF features.
Polymers 15 02954 g008
Figure 9. Partial dependency graph of the number of (a) methyl groups and (b) carbonyl groups.
Figure 9. Partial dependency graph of the number of (a) methyl groups and (b) carbonyl groups.
Polymers 15 02954 g009
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Huang, D.; Li, Z.; Wang, K.; Zhou, H.; Zhao, X.; Peng, X.; Zhang, R.; Wu, J.; Liang, J.; Zhao, L. Probing the Effect of Photovoltaic Material on Voc in Ternary Polymer Solar Cells with Non-Fullerene Acceptors by Machine Learning. Polymers 2023, 15, 2954. https://doi.org/10.3390/polym15132954

AMA Style

Huang D, Li Z, Wang K, Zhou H, Zhao X, Peng X, Zhang R, Wu J, Liang J, Zhao L. Probing the Effect of Photovoltaic Material on Voc in Ternary Polymer Solar Cells with Non-Fullerene Acceptors by Machine Learning. Polymers. 2023; 15(13):2954. https://doi.org/10.3390/polym15132954

Chicago/Turabian Style

Huang, Di, Zhennan Li, Kuo Wang, Haixin Zhou, Xiaojie Zhao, Xinyu Peng, Rui Zhang, Jipeng Wu, Jiaojiao Liang, and Ling Zhao. 2023. "Probing the Effect of Photovoltaic Material on Voc in Ternary Polymer Solar Cells with Non-Fullerene Acceptors by Machine Learning" Polymers 15, no. 13: 2954. https://doi.org/10.3390/polym15132954

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop