Machine Learning for Perovskite Solar Cells and Component Materials: Key Technologies and Prospects

Probably not a substantial claim: may not pass S3 (load-bearing).

The test set cannot replace the validation set to perform hyperparameter search because doing so leads to overfitting and makes the test set lose its ability to evaluate the generalization capability of the model.

“The test set cannot replace the validation set to perform hyperparameter search because this would lead to overfitting and make the test set lose its ability to evaluate the generalization capability of the model.[148]”

• cites [148] 148. P. BURMAN, Biometrika 1989, 76, 503. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Interpretation and explanations are not uniformly defined and distinguished.

“Where interpretation and explanations are not uniformly defined and distinguished.[163,164]”

• cites [163] 163. L. H. Gilpin, D. Bau, B. Z. Yuan, A. Bajwa, M. Specter, L. Kagal, IEEE 5th Int. Conf. on Data Science and Advanced Analytics 2018, 80. https://doi.org/10.1109/dsaa.2018.00018. sourceSupported
• cites [164] 164. J.-X. Mi, A.-D. Li, L.-F. Zhou, IEEE Access 2020, 8, 191969.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Common interpretable algorithmic models (GLM, DT, SR) have a similar characteristic that the inner operating mechanism is intuitive, understandable and transparent.

“They have a similar characteristic that the inner operating mechanism of the models is intuitive, understandable and transparent.[170,171]”

• cites [170] 170. S. Zheng, C. Ding, Pattern Recogn. Lett. 2020, 131, 227. sourcePartly supported

  In “Decision Theory Meets Explainable AI” (Background > Decision Theory; page unavailable), the paper says linear models make importance and utility easy to define, and its Explainable Artificial Intelligence background describes rules as an interpretable form of knowledge. It does not identify GLMs, decision trees, and symbolic regression as a group or say that all three have intuitive, transparent inner mechanisms; it also notes explainability challenges for rule-based systems. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [171] 171. Y. Wang, N. Wagner, J. M. Rondinelli, MRS Commun. 2019, 9, 793.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Applications of ML in perovskite solar cells and component materials have only begun to boom in the last two years.

“However, the applications of ML in perovskite solar cells and component materials has only begun to boom in the last two years”

Probably not a substantial claim: may not pass S3 (load-bearing).

The power conversion efficiency of perovskite solar cells has improved to 25.7% in less than 10 years.

“the power conversion efficiency (PCE) has improved to 25.7% in <10 years.[31]”

• cites [31] 31. D.-K. Lee, N.-G. Park, Sol. RRL 2021, 6, 2100455.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims. The quote only partly states this claim.

In organic solar cell donor materials, average model accuracy is lower when SMILES strings are used as input, while average model accuracy is higher when fingerprints of length greater than 1000 bits are used as input.

“Among them, the average accuracy of the model is lower when SMILES strings are used as input, while the average accuracy of the model is higher when fingerprints of length >1000 bits are used as input.”

• cites [109] 109. A. Mauri, V. Consonni, M. Pavan, R. Todeschini, MATCH Commun. Math. Comput. Chem. 2006, 56, 237.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Data collection and regularization for PSCCM are difficult, accounting for >70% of the entire model building process and requiring extensive knowledge.

“Third, data collection and regularization are difficult.In our experience, data collection and regularization are not only extremely heavy, accounting for >70% of the entire model building process, but also require the collectors to have extensive knowledge and background related to PSCCM.”

Probably not a substantial claim: may not pass S4 (based).

At this stage, the physicochemical and structural data of perovskite materials, the component and performance test data of device-level PSCs are still in a barren and isolated state.

“At this stage, the physicochemical and structural data of perovskite materials, the component and performance test data of device-level PSCs are still in a barren and isolated state.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The articles that combine ML with PSCCM exploded in the last two years.

“First, the articles that combine ML with PSCCM exploded in the last two years”

Probably not a substantial claim: may not pass S4 (based).

Predicting physicochemical properties and screening component materials related to perovskite solar cells is a strong point of ML.

“Predicting the physicochemical properties and screening the component materials related to perovskite solar cells is the strong point of ML.”

Probably not a substantial claim: may not pass S4 (based).

ML can efficiently handle nonlinearities and game-playing problems in complex systems composed of multiple materials, beyond traditional computational and simulation methods.

“Its ability and efficiency to handle nonlinear and game-playing problems is unmatched by traditional simulation computing software and trial-error experiments.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Data normalization can increase efficiency and accuracy of the model when order of magnitude differences are too large.

“when the order of magnitude differences of data are too large, data normalization can be used, which can increase the efficiency and accuracy of the model.[75]”

• cites [75] 75. Y. Liu, W. Yan, H. Zhu, Y. Tu, L. Guan, X. Tan, Org. Electron. 2022, 101, 106426.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

ML has been deeply wedged into the materials field.

“As of today, ML has also been deeply wedged into the materials field.[14][15][16][17][18][19][20][21]”

• cites [14] 14. G. R. Schleder, A. C. M. Padilha, C. M. Acosta, M. Costa, A. Fazzio, J. Phys.: Mater. 2019, 2, 032001.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [15] 15. G. H. Gu, J. Noh, I. Kim, Y. Jung, J. Mater. Chem. A 2019, 7, 17096.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [16] 16. G. L. W. Hart, T. Mueller, C. Toher, S. Curtarolo, Nat. Rev. Mater. 2021, 6, 730.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

The computational cost of ML is much lower than that of contemporaneous experimental and computational methods.

“The computational cost of ML is much lower than that of contemporaneous experimental and computational methods.”

May combine two claims.

ML can predict properties of unknown perovskite solar cells from existing data, accelerating design and development of perovskite solar cells.

“ML can predict the properties of unknown PSCs from existing data of PSCs, which can accelerate the design and development of PSCs.[55]”

• cites [55] 55. W. Xu, Z. Liu, R. T. Piper, J. W. P. Hsu, Sol. Energy Mater. Sol. Cells 2023, 249, 112055.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

ML accelerates the process of developing new materials and brings inspirations and new research perspectives to materials scientists.

“ML, the greatest product of the data-driven epoch, accelerates the process of developing new materials and even bring inspirations that are beyond the reach of human hands and new research perspectives to materials scientists.[26][27][28][29]”

• cites [26] 26. Y. Juan, Y. Dai, Y. Yang, J. Zhang, J. Mater. Sci. Technol. 2021, 79, 178.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [27] 27. P. Raccuglia, K. C. Elbert, P. D. F. Adler, C. Falk, M. B. Wenny, A. Mollo, M. Zeller, S. A. Friedler, J. Schrier, A. J. Norquist, Nature 2016, 533, 73.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [28] 28. C. Gao, X. Min, M. Fang, T. Tao, X. Zheng, Y. Liu, X. Wu, Z. Huang, Adv. Funct. Mater. 2021, 32, 2108044.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Most data in the perovskite solar cells and component materials field are tabular data, with no obvious space or time links between features, and sample volume is small and heterogeneous, leading to ensemble learning algorithms with trees typically outperforming deep learning.

“Most of the data in the field of PSCCM are tabular data.There are usually no obvious space or time links between features, and the sample volume of the data is small and heterogeneous, which leads to ensemble learning algorithms with trees that typically outperform DL. [122,123]”

• cites [122] 122. R. Shwartz-Ziv, A. Armon, Inform. Fusion 2022, 81, 84.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [123] 123. O. Sagi, L. Rokach, WIREs Data Mining Knowl. Discov. 2018, 8, e1249.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep learning belongs to a subset of ML, originated from artificial neural networks, and excels at processing unstructured data such as images, text, and sound.

“DL belongs to a subset of ML that originated from artificial neural networks and excels to process unstructured data (images, text and sound).[128,129]”

• cites [128] 128. R. Miotto, F. Wang, S. Wang, X. Jiang, J. T. Dudley, Briefings in Bioinformatics 2018, 19, 1236.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [129] 129. A. Shrestha, A. Mahmood, IEEE Access 2019, 7, 53040.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

An appropriate amount of experimental and computational data on perovskite solar cells and component materials already exists, and related ML technologies are mature.

“An appropriate amount of experimental and computational data on perovskite solar cells and component materials (PSCCM) already exists, and the related ML technologies are mature.”

May combine two claims.

ML can be applied only when a large amount of perovskite solar cell related data exists, supporting the importance of data for ML models.

“This illustrates that the ML can be applied only when a large amounts of PSCs related data exists, and further supports the importance of the data for the ML models.”

May combine two claims.

The lower accuracy with SMILES may be because SMILES string is close to the original information causing some ML algorithms not to extract hidden features, while longer fingerprints contain more chemical information making them more unique and easily available for model training.

“This may be due to the fact that the SMILES string is still close to the original information of the data that causes some ML algorithms not to extract hidden features, while longer fingerprints can contain more chemical information making them more unique and easily available for model training.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Proper extraction of features largely affects the simplicity and generalization capability of the model and the effectiveness of model interpretation.

“The proper extraction of features largely affects the simplicity and generalization capability of the model and the effectiveness of the model interpretation.[78][79][80]”

• cites [78] 78. V. Bolón-Canedo, A. Alonso-Betanzos, Inform. Fusion 2019, 52, 1.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [79] 79. P. M. Khan, K. Roy, Expert Opinion on Drug Discovery 2018, 13, 1075.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [80] 80. Y. Li, T. Li, H. Liu, Knowledge and Information Systems 2017, 53, 551.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

SMILES string and chemical structure of a molecule are in unique correspondence that can be used as a fingerprint to express differences of chemical structure, elemental and linkage properties.

“The SMILES string and the chemical structure of the molecule are in unique correspondence that can be used as a fingerprint of the chemical structure to express differences of chemical structure, elemental and linkage properties, etc. [107]”

• cites [107] 107. Z.-W. Zhao, M. del Cueto, Y. Geng, A. Troisi, Chem. Mater. 2020, 32, 7777.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Data is the life of ML models, and data is more important than the algorithm itself.

“It is no exaggeration that data is the life of the ML models, and even data is more important than the algorithm itself.[57]”

• cites [57] 57. C. Chen, Y. Zuo, W. Ye, X. Li, Z. Deng, S. P. Ong, A Adv. Energy Mater. 2020, 10, 1903242.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

It is difficult to conduct a comprehensive analysis of the complicated influencing factors of perovskite solar cells by a low-cost experimental or computational method.

“For the above-mentioned complicated influencing factors, it is difficult to conduct a comprehensive analysis by a low-cost experimental or computational method.”

May combine two claims.

Ensemble learning avoids shortcomings inherent in a single model or a model with a single set of parameters, and the final optimized prediction model has strong robustness and anti-overfitting capability.

“It avoids the shortcomings inherent in a single model or a model with a single set of parameters, and the final optimized prediction model has strong robustness and anti-overfitting capability.[124,125]”

• cites [124] 124. X. Dong, Z. Yu, W. Cao, Y. Shi, Q. Ma, Front. Computer Sci. 2019, 14, 241.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [125] 125. Y. Ren, L. Zhang, P. N. Suganthan, IEEE Comput. Intell. Mag. 2016, 11, 41.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

ML is subtly replacing traditional techniques in numerous fields such as biopharmaceuticals, image identification, and data mining.

“ML is subtly replacing traditional techniques in numerous fields, such as, biopharmaceuticals, [11] image identification, [12] data mining, [13] etc.”

• cites [11] 11. Y. Yang, R. Zhang, Z. Li, L. Mei, S. Wan, H. Ding, Z. Chen, J. Xing, H. Feng, J. Han, H. Jiang, M. Zheng, C. Luo, B. Zhou, J. Med. Chem. 2020, 63, 1337.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [12] 12. A. Esteva, K. Chou, S. Yeung, N. Naik, A. Madani, A. Mottaghi, Y. Liu, E. Topol, J. Dean, R. Socher, npj Digital Med. 2021, 4, 5.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [13] 13. E. A. Olivetti, J. M. Cole, E. Kim, O. Kononova, G. Ceder, T. Y.-J. Han, A. M. Hiszpanski, Appl. Phys. Rev. 2020, 7, 041317.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Dimensionality reduction of features can enhance model robustness, reduce computational cost, and remove noise, but may lose important information and obscure feature meaning; it is mostly used for image-type data.

“The dimensionality reduction of features can enhance the robustness of the model, reduce the computational cost, and remove the noise in the dataset, but this may lose some important information and obscure the meaning of the features, which are mostly used in processing image-type data.[89,90]”

• cites [89] 89. M. Lualdi, M. Fasano, J. Proteomics 2019, 198, 18.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [90] 90. T. Zhou, Z. Song, K. Sundmacher, Engineering 2019, 5, 1017.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Molecular fingerprint is now the most commonly used technique for transformation of molecular descriptors because it maximally preserves structural and physicochemical information of the molecule as input to ML models.

“Molecular fingerprint is now the most commonly used technique for the transformation of molecular descriptors, because it maximally preserves the structural and physicochemical information of the molecule as an input to the ML models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Perovskite solar cells are the third-generation solar cell with the most promising commercialization.

“It is the third-generation solar cell with the most promising commercialization, [30]”

• cites [30] 30. X. Luo, X. Lin, F. Gao, Y. Zhao, X. Li, L. Zhan, Z. Qiu, J. Wang, C. Chen, L. Meng, X. Gao, Y. Zhang, Z. Huang, R. Fan, H. Liu, Y. Chen, X. Ren, J. Tang, C.-H. Chen, D. Yang, Sci. China: Chem. 2022, 65, 2369.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The choice of suitability algorithm is usually judged by the credibility level on the model, the type and volume of data.

“The choice of suitability algorithm is usually judged by the credibility level on the model, the type (table, image, text, etc.) and volume of data.”

Probably not a substantial claim: may not pass S2 (checkable).

Intervention of ML for the study of perovskite solar cells may be another promising direction.

“Therefore, the intervention of ML for the study of PSCs may be another promising direction.[53,54]”

• cites [53] 53. Z. Wang, M. Yang, X. Xie, C. Yu, Q. Jiang, M. Huang, H. Algadi, Z. Guo, H. Zhang, Adv. Composites Hybrid Mater. 2022, 5, 2700.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [54] 54. W. Yan, Y. Liu, Y. Zang, J. Cheng, Y. Wang, L. Chu, X. Tan, L. Liu, P. Zhou, W. Li, Z. Zhong, Nano Energy 2022, 99, 107394.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

LIME provides a local explanation for the black box model, explaining the prediction results of a single sample.

“LIME provides a local explanation for the black box model, that is, it explains the prediction results of a single sample.”

• cites [176] 176. S. Lundberg, S.-I. Lee, Adv. Neural Inform. Processing Syst. 2017, 30, 4765.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

The performance of PSCs is influenced by multiple scales and multiple domains of physicochemical features.

“The performance of the PSCs is influenced by multiple scales and multiple domains of physicochemical features, such as the properties of the material itself, the compatibility between materials, and the structure and fabrication of the PSCs”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Traditional DFT calculation and simulation methods are no longer applicable in terms of efficiency and performance for PSCs, whereas ML is the best solution nowadays.

“At this point, the traditional DFT calculation and simulation methods are no longer applicable in terms of efficiency and performance, whereas ML is the best solution nowadays.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

The concepts related to model validation and optimization are easily confused, and the application of these techniques is not standardized.

“The concepts related to model validation and optimization are easily confused, and the application of these techniques is not standardized”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Optimization of hyperparameters is essential because hyperparameters under the same algorithm can directly affect the performance of the model, which can lead to underfitting or overfitting.

“Optimization of hyperparameters is essential because hyperparameters under the same algorithm can directly affect the performance of the model, which can lead to underfitting or overfitting of the model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SHAP has developed several different explainers for black box models, among which TreeExplainer and DeepExplainer are used for ensemble learning and DL respectively.

“SHAP has developed several different explainers for black box models.Among them, TreeExplainer and DeepExplainer are used for ensemble learning and DL respectively.[122,178]”

• cites [122] 122. R. Shwartz-Ziv, A. Armon, Inform. Fusion 2022, 81, 84.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [178] 178. J. Klaise, A. Looveren, G. Vacanti, J. Mach. Learn. Res. 2021, 22, 8194.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The FA (formamidinium) content plays the most important contribution to the predicted value of PCE, and high FA content contributes more positively to PCE than the negative contribution when FA content is low.

“Among them, the FA (formamidinium) content plays the most important contribution to the predicted value of PCE, and the high FA content contributes more positively to PCE than the negative contribution when the FA content is low.”

• cites [205] 205. A. Mannodi-Kanakkithodi, M. K. Y. Chan, J. Mater. Sci. 2022, 57, 10736.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In a sample, the model prediction for PCE is 23.53%, and the SHAP explanation shows that the high prediction value is mainly attributed to high FA content, high hole mobility and suitable bandgap.

“In this sample, the model prediction for PCE is 23.53%, and the SHAP explanation shows that the high prediction value is mainly attributed to the high FA content, high hole mobility and suitable bandgap.”

• cites [205] 205. A. Mannodi-Kanakkithodi, M. K. Y. Chan, J. Mater. Sci. 2022, 57, 10736.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

For PSC, the role of model interpretation may be more important than model performance, because it can give materials scientists guidance or inspiration to explore feature connections and potential causal laws behind complex physicochemical systems.

“For PSC, the role of model interpretation may be more important than model performance, because it can give materials scientists some guidance or inspiration to explore the feature connections and potential causal laws behind complex physicochemical systems.”

May combine two claims.

ML can greatly improve the efficiency of material screening and reduce the computational burden of DFT.

“Of course, ML can greatly improve the efficiency of material screening and reduce the computational burden of DFT, which demonstrates the great potential of combining ML and DFT in the field of perovskite materials.”

Probably not a substantial claim: may not pass S3 (load-bearing).

LIME can be used for the explanation of any black box model and input data.

“LIME can be used for the explanation of any black box model and input data, but it also has some drawbacks.”

• cites [176] 176. S. Lundberg, S.-I. Lee, Adv. Neural Inform. Processing Syst. 2017, 30, 4765.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

k-fold CV has a smaller error and is more representative of the total sample than Hold-out.

“This has a smaller error and more representative of the total sample.[153]”

• cites [153] 153. D. Krstajic, L. J. Buturovic, D. E. Leahy, S. Thomas, J. Cheminform. 2014, 6, 10.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Performing repeated LIME operations may result in different explanations due to differences in generated new samples.

“Performing repeated operations may result in different explanations due to the differences of generated new samples.”

May combine two claims.

The performance of ensemble learning and DL is usually better than classical learning, but their internal working mechanisms also become more complex.

“The performance of ensemble learning and DL is usually better than classical learning, but their internal working mechanisms also become more complex (e.g., DL uses many nonlinear overlays to improve the performance of the model). [157]”

• cites [157] 157. P. P. Angelov, E. A. Soares, R. Jiang, N. I. Arnold, P. M. Atkinson, WIREs Data Mining Knowl. Discov. 2021, 11, e1424.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Hold-out (Model Validation) does not have model optimization ability because it only splits data into training and test sets and performs training and testing once.

“It only splits the data into training set and test set without setting the validation set, and only performs model training and testing once, so it does not have model optimization ability.”

Probably not a substantial claim: may not pass S3 (load-bearing).

LIME ignores the relationship between features in the generation of new samples, which may lead to the generation of some impossible samples.

“For example, it ignores the relationship between features in the generation of new samples, which may lead to the generation of some impossible samples.”

May combine two claims.

Traditional methods for studying perovskite properties (simple descriptors or DFT calculations) are either less accurate or require a lot of computational resources and time.

“However, these traditional methods are either less accurate or require a lot of computational resources and time.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The interpretable algorithmic models generally belong to the global interpretation type, able to interpret the entire model's operating mechanism based on the complete data set.

“The interpretable algorithmic models generally belong to the global interpretation type, namely, they are able to interpret the entire model's operating mechanism based on the complete data set.[167]”

• cites [167] 167. M. R. Zafar, N. M. Khan, arXiv preprint arXiv:1906.10263, 2019. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The perovskite solar cells and component materials field commonly uses ML to deal with classification and regression problems.

“The field of PSCCM commonly uses ML to deal with classification and regression problems”

Classified as probably not a substantial claim. May combine two claims.

The 67 papers with detailed machine learning technical descriptions and representation are included in Table 7 and deconstructed in 10 dimensions.

“The 67 papers with detailed ML technical descriptions and representation are included in Table 7 and deconstructed in 10 dimensions”

May combine two claims.

The application of ML models in perovskite materials is mostly as high-throughput virtual materials screening, which greatly reduces the credibility and practicability of ML models.

“But it also leads to the application of ML models in perovskite materials mostly as high-throughput virtual materials screening, which greatly reduces the credibility and practicability of ML models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The theory of SHAP explanation is to calculate the cumulative contribution of the features to the predicted values.

“The theory of SHAP explanation is to calculate the cumulative contribution of the features to the predicted values.”

• cites [177] 177. H. Chen, S. Lundberg, S.-I. Lee, in Explainable AI in Healthcare and Medicine, (Eds: A. Shaban-Nejad, M. Michalowski, D. L. Buck- eridge), Springer, Cham, Switzerland 2020, 914, 261.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Liu et al. used SHAP to explain the model PCE_RF.

“After that, they used SHAP to explain the model PCE RF .”

• cites [205] 205. A. Mannodi-Kanakkithodi, M. K. Y. Chan, J. Mater. Sci. 2022, 57, 10736.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The parameters of PSCs are easily verified experimentally, which increases the credibility of ML model for application in PSCs.

“Compared the screening of virtual perovskite materials by ML model, the parameters of PSCs are easily verified experimentally, which increases the credibility of ML model for application in PSCs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SHAP visualizes the importance ranking of the 13 features and the contribution of each feature.

“As shown in Figure 9a, SHAP visualizes the importance ranking of the 13 features and the contribution of each feature.”

• cites [205] 205. A. Mannodi-Kanakkithodi, M. K. Y. Chan, J. Mater. Sci. 2022, 57, 10736.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

SR is a representative interpretable ML algorithm that attempts to create optimal mathematical formulations to map the connection between features and predictive targets.

“SR is a representative interpretable ML algorithm that attempts to create optimal mathematical formulations to map the connection between features and predictive targets.[172,173]”

• cites [172] 172. D. A. Augusto, H. J. C. Barbosa, Proc. Sixth Brazilian Symposium on Neural Networks 2000, 1, 173.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [173] 173. G. Peake, J. Wang, Proc. 24th ACM SIGKDD Int. Conf. Knowledge Discovery & Data Mining 2018, 19.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The interpretation of the model is used to determine whether the model captures valuable features and thus judge whether the model can be transferred to the prediction of other samples, which is significant for small sample databases.

“The interpretation of the model is used to determine whether the model captures valuable features and thus judge whether the model can be transferred to the prediction of other samples, which is significant for small sample databases.”

Probably not a substantial claim: may not pass S3 (load-bearing).

There are two main sources of data related to PSC: manual experiments or literature collection, and generation by simulation software.

“There are two main sources of data related to PSC, one is manual experiments or literature collection, and the other is generation by simulation software.”

Probably not a substantial claim: may not pass S4 (based).

The massive accumulation of data related to perovskite solar cells and component materials creates space for machine learning to cut in.

“Thanks to the massive accumulation of data related to PSCCM, there is space for ML to cut in.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SHAP is the most popular explanation method for black box models nowadays, originated from game theory, and was first applied to local and global explanation of black box models in 2017.

“SHAP is the most popular explanation method for black box models nowadays, which originated from game theory and was first applied to local and global explanation of black box models in 2017.[177]”

• cites [177] 177. H. Chen, S. Lundberg, S.-I. Lee, in Explainable AI in Healthcare and Medicine, (Eds: A. Shaban-Nejad, M. Michalowski, D. L. Buck- eridge), Springer, Cham, Switzerland 2020, 914, 261.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The operational mechanism of the black box model is in urgent need of interpretation, because it can promote credibility of the models and optimization of features and models.

“The operational mechanism of the”

• cites [160] 160. M. Sahakyan, Z. Aung, T. Rahwan, IEEE Access 2021, 9, 135392.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: Model performance is only the result of model decisions, and the reasons for model decisions are sometimes not clear to us.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

If the operating mechanism of the model is difficult to understand, it will lead to credibility problems, especially for screening virtual perovskite materials and designing new materials.

“If the operating mechanism of the model is difficult to understand, it will lead to credibility problems, especially the screening of virtual perovskite materials and the design of new materials.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In contrast to traditional ML algorithms, deep learning can automatically extract efficient features from unstructured data and train them.

“In contrast to traditional ML algorithms, it can automatically extract efficient features from unstructured data and train them.[130,131]”

• cites [130] 130. S. So, T. Badloe, J. Noh, J. Bravo-Abad, J. Rho, Nanophotonics 2020, 9, 1041.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [131] 131. G. P. Zhang, IEEE Trans. Systems, Man Cybernetics, Part c 2000, 30, 451.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

The features of these materials involve thousands of properties such as physical, electronic, structural, morphological, and chemical etc.

“The features of these materials in turn involve thousands of properties such as physical, electronic, structural, morphological, and chemical etc.”

Classified as probably not a substantial claim.

The model building process has a vague model training process, mainly shown by the lack of evaluation of model underfitting and overfitting.

“4) The vague model training process, mainly shown that the lack of evaluation of model underfitting and overfitting.”

Classified as probably not a substantial claim. May combine two claims.

ML models composed of ensemble learning or DL algorithms generally have better performance for PSCCM, but these complex models are less interpretable, which makes them less credible.

“ML models composed of ensembles learning or DL algorithms generally have better performance for PSCCM, but these complex models are less interpretable, which makes them less credible.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

These functional relations may be very important because it may reveal potential physicochemical relations between microscopic features and the attributes of macroscopic target.

“These functional relations may be very important because it may reveal potential physicochemical relations between microscopic features and the attributes of macroscopic target.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

This problem can be avoided using ML replacement, because the data comes from multiple experimental data and the model will analyze and compare to make the most representative judgment.

“But this problem can be avoided using ML replacement, because the data comes from multiple experimental data and the model will analyze and compare to make the most representative judgment.”

May combine two claims.

The accuracy of the ML model for classifying recombination processes was 82%, higher than 70% when using the ideal factor.

“The accuracy of the ML model was 82% higher than 70% when using the ideal factor.”

• cites [228] 228. A. Talapatra, B. P. Uberuaga, C. R. Stanek, G. Pilania, Chem. Mater. 2021, 33, 845.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Methods that help select the number of inputs have limitations, such as not taking into account potential physicochemical links between features and not being able to select the types of features.

“Although there are methods that can help to select the number of inputs (see Section 2.2.1), these methods also have limitations, such as not taking into account the potential physicochemical links between features and not being able to select the types of features.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The model building process has a vague hyperparameter search range.

“5) The vague hyperparameter search range.”

Probably not a substantial claim: may not pass S4 (based).

The model building process has an overly sparse and unbalanced distribution of training data.

“1) Overly sparse and unbalanced distribution of training data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In PSCs, ML is mainly used for feature analysis to find key features that affect cell properties and to accelerate and guide experimental process, requiring complete feature construction logic and high interpretability.

“In PSCs, ML is mainly used for feature analysis to find the key features that affect the cell properties and to accelerate and guide the experimental process.Of course, it requires complete feature construction logic and high interpretability of the model.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Low cost-performance models are characterized by poor generalization capability.

“2) The generalization capability of the model is poor.”

May combine two claims.

DFT is an important method of generating computational data for perovskite materials, but its computational cost for high-precision data is too high.

“DFT is an important method of generating computational data for perovskite materials, but its computational cost for high-precision data is too high.”

Probably not a substantial claim: may not pass S4 (based).

There exist various ML algorithms that can generate functional relations between features and target attributes only from data.

“There exist various ML algorithms that can the generate functional relations between features and target attributes only from data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Low cost-performance models are characterized by the difficulty of obtaining inputs not being less than the difficulty of obtaining labels.

“1) The difficulty of obtaining the inputs is not less than the difficulty of obtaining the labels of the model.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The study of ML application in PSCs has important reference value and transfer possibility for other spaces (LED, photodetectors, etc.).

“PSCs are one of the complex physicochemical spaces composed of perovskite materials, the study of ML application in PSCs has important reference value and transfer possibility for other spaces (LED, photodetectors, etc.).”

May combine two claims.

The supplementation of PSCCM data is a long-term project, and it is likely that some of the data are already inadequate (e.g., the information of tandem or large-area PSCs, information of non-toxic perovskite materials, etc.).

“The supplementation of PSCCM data is a long-term project, and it is likely that some of the data are already inadequate (e.g., the information of tandem or large-area PSCs, information of non-toxic perovskite materials, etc.).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Bayesian optimization is an active learning algorithm that allows iterative optimization of the model based on small samples, thus reducing the reliance on high-cost experimental data of PSCCM.

“Bayesian optimization is an active learning algorithm that allows iterative optimization of the model based on small samples, thus reducing the reliance on high-cost experimental data of PSCCM.”

May combine two claims.

The molecular descriptors generated by currently used feature transformation tools do not represent the full properties of the material, and cannot take into account the requirements of accuracy, speed, and interpretability.

“The molecular descriptors generated by the currently used feature transformation tools (see Section 2.2.2) do not represent the full properties of the material, and cannot take into account the requirements of accuracy, speed, and interpretability.[255]”

• cites [255] 255. A. Hernandez, A. Balasubramanian, F. Yuan, S. A. M. Mason, T. Mueller, npj Comput. Mater. 2019, 5, 112.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Low cost-performance models are characterized by predicted target properties that can be expressed by a simple function mapping.

“4) The predicted target properties can be expressed by a simple function mapping.”

May combine two claims.

Characterization tests usually require the fabrication of complete devices and the devices used in different characterizations are not the same, which results in a waste of experimental resources.

“First, these characterization tests usually require the fabrication of complete devices and the devices used in different characterizations are not the same, which results in a waste of experimental resources.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The model building process has a shortage and homogeneity of model interpretation.

“7) The shortage and homogeneity of model interpretation.”

Probably not a substantial claim: may not pass S4 (based).

The model building process has an unreasonable evaluation method for model generalization capability, mainly shown by the unreasonable use of training set, validation set and test set.

“6) Unreasonable evaluation method for model generalization capability, mainly shown that the unreasonable use of training set, validation set and test set.”

Probably not a substantial claim: may not pass S4 (based).

The model building process has unreasonable splitting of training and test sets.

“3) Unreasonable splitting of training and test sets.”

May combine two claims.

The recombination processes in PSCs are classified into three categories (band-to-band, grain boundary traps, and interfacial traps (Front/Back interface)), which cannot be effectively classified when the conventional ideal factor is used.

“As shown in Figure 15e, the recombination processes in PSCs are classified into three categories (band-to-band, grain boundary traps, and interfacial traps (Front/Back interface)), which cannot be effectively classified when the conventional ideal factor is used to classify them.”

• cites [228] 228. A. Talapatra, B. P. Uberuaga, C. R. Stanek, G. Pilania, Chem. Mater. 2021, 33, 845.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

PSCs are complex physicochemical systems composed of a variety of materials.

“PSCs are complex physicochemical systems composed of a variety of materials, including perovskite materials, organic small molecule materials, polymer materials, metal oxides and many other materials with different properties.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Feature selection requires the researcher to have knowledge of both PSCCM and ML fields.

“which requires the researcher to have knowledge of both PSCCM and ML fields.”

May combine two claims.

The autonomous robotic experimental flow allows for uninterrupted and long experiments without human involvement, and analyze data via ML to optimize the experiments, which brings the possibility of efficient and costeffective experiments.

“In contrast to traditional experimental methods, the autonomous robotic experimental flow allows for uninterrupted and long experiments without human involvement, and analyze data via ML to optimize the experiments, which brings the possibility of efficient and costeffective experiments.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

There is no essential difference between meta-learning and transfer learning.

“There is no essential difference between meta-learning and transfer learning.”

May combine two claims.

The PCE, Voc, Jsc, and FF are closely related for the same PSC, so the knowledge learned in the PCE prediction model can be transferred to the prediction models of Voc, Jsc, and FF.

“Specifically, the PCE, Voc, Jsc, and FF are closely related for the same PSC, so the knowledge learned in the PCE prediction model can be transferred to the prediction models of Voc, Jsc, and FF.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The device-level characterization of PSCs includes a variety of tests, such as electrical performance test, External quantum efficiency (EQE) test, EIS test and stability test.

“The device-level characterization of PSCs includes a variety of tests, such as electrical performance test, External quantum efficiency (EQE) test, EIS test and stability test.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In perovskite materials, ML applications are tilted toward screening of virtual materials, requiring high generalization capability and less demanding interpretation.

“In perovskite materials, the applications of ML are tilted toward the screening of virtual materials, which requires the model to have high generalization capability, and the interpretation and feature analysis of the model are less demanding.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

ML generation of new component materials of PSC is the deconstruction and reconstruction of known materials, which in turn generates new materials that better meet the target requirements.

“It is the deconstruction and reconstruction of known materials, which in turn generates new materials that better meet the target requirements.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

For electrical performance characterization of PSCs with popular components and structures, test results may not be the same or even test errors may occur due to small differences in each experimenter's skill level and characterization equipment.

“For example, for the electrical performance characterization of PSCs with popular components and structures, the test results may not be the same or even test errors may occur due to small differences in each experimenter's skill level and characterization equipment.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Low cost-performance models are characterized by requiring a large amount of data for fitting.

“3) The fitting of the model requires a large amount of data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

It is possible to transfer the inputs used under the same label for similar perovskite materials.

“Similarly, it is possible to transfer the inputs used under the same label for similar perovskite materials.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Adequate and abundant data of PSCCM is a guarantee for the application of ML into the following two stages.

“Adequate and abundant data of PSCCM is a guarantee for the application of ML into the following two stages.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Device-level characterization tests are essentially a mapping between device performance and the properties of the material used, which is close to the concept of ML.

“Device-level characterization tests are essentially a mapping between device performance and the properties of the material used, which is close to the concept of ML.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Due to the short time of combination between ML and PSCCM, many study steps are incomplete or even unreasonable.

“Due to the short time of combination between ML and PSCCM, many study steps are incomplete or even unreasonable.”

• AssumptionRests on an assumption: The combination between ML and PSCCM has been short.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Unlike high-throughput material screening via ML, ML generation of new component materials of PSC is more purposeful and efficient.

“Unlike high-throughput material screening via ML, ML generation of new component materials of PSC is more purposeful and efficient.”

Classified as probably not a substantial claim. May combine two claims.

The functional relations between the performance of PSCs and the properties of the component materials are used to help people fully understand the performance of PSCs, and to screen and design new component materials that meet the requirements.

“In other words, the functional relations between the performance of PSCs and the properties of the component materials are used to help people fully understand the performance of PSCs, and to screen and design new component materials that meet the requirements.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

When target properties are classified more elaborately, the performance of models of that kind will be better, but when models are used for different kinds of targets, the performance will be greatly reduced.

“When the target properties are classified more elaborately, the performance of models of that kind will be better, but when the models are used for different kinds of targets, the performance will be greatly reduced.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The model building process has overly redundant algorithm comparisons.

“2) Overly redundant algorithm comparisons.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

For PSC, ML can generate new passivation/interface materials composed of organic small molecules based on lewis acid-base interactions, and additive materials that improve the growth of perovskite crystals can also be generated based on structural information such as functional groups and chain length.

“For PSC, ML can generate new passivation/interface materials composed of organic small molecules based on lewis acid-base interactions, and additive materials that improve the growth of perovskite crystals can also be generated based on structural information such as functional groups and chain length.”