Machine learning facilitates the screening of interface passivation materials for perovskite solar cells. Perovskite solar cells have become a shining new star in the field of photovoltaics due to their excellent photoelectric performance and low manufacturing cost. up to 25.7%. However, its efficiency, stability, and large-area preparation process still need to be further optimized to shorten the distance between its theoretical limit of performance and commercial application.In recent years, low-dimensional modification of the interface of perovskite solar cells has been widely used to improve the efficiency and stability of devices. However, the materials used for low-dimensional interface processing are various and complex in structure, and it usually requires a lot of repeated experiments to select a suitable passivation material. This trial-and-error method of "fried dishes" is costly, time-consuming, and is easily affected by the subjective judgment of the experimenter or the fluctuation of objective experimental conditions; at present, there is no exact interface passivation material screening criteria as a new passivation method. Reference for material development. Therefore, this work adopts the research method of machine learning to study the relationship between the molecular feature descriptors of interface passivation materials and the energy conversion efficiency. Materials screening with the assistance of machine learning models can help researchers interpret key molecular features of perovskite interfacial passivation, quantify screening criteria, and predict candidate materials. The machine learning model used in this work is expected to be applied to a wider range of organic molecular screening to improve the performance and stability of optoelectronic devices.
Figure 1. Schematic diagram of thesis research ideas
【Introduction】
Recently, the team of professors Liu Zhe, Li Zhen and Wang Hongqiang from the Nano Energy Materials Research Center of Northwestern Polytechnical University published a research titled "Machine-Learning-Assisted Screening of Interface Passivation Materials for Perovskite Solar Cells" in the internationally renowned journal ACS Energy Letters paper. In this work, the molecular characteristics of 19 groups of interface passivation materials and the improvement ratio of device efficiency after treatment were modeled and trained by machine learning methods, and the importance ranking of molecular characteristics of passivation materials was obtained and the quantification of high-efficiency passivation materials was obtained. screening criteria. On this basis, according to the prediction results of the model, several better materials were selected for experimental verification, and finally the device efficiencies of 22.36% and 24.47% were obtained in FAMACs-based and FAMA-based perovskite solar cell devices, respectively.
【Article Highlights】
Point 1: Data preprocessing - increasing data diversity and removing invalid descriptors
Through principal element analysis (PCA), various features of the material are reduced in dimension to realize the visual analysis of material feature distribution. On this basis, the Latin Hypercube Sampling (LHS) method is used to select passivation materials distributed in different intervals, and then supplement the original data set to enhance the diversity of the initial data set. In addition, the Pearson correlation coefficient between the feature descriptors of the material was fitted to remove molecular features with strong correlation and improve the accuracy and interpretability of the machine learning model.
Figure 2. Principal component analysis results and Pearson correlation of material characteristics
Point 2: Model comparison - using integrated learning to combine the advantages of multiple regression algorithms
This work uses a variety of regression algorithms to train the data set, including random forests, gradient boosting, support vector machines, neural networks, and Gaussian processes and other regression algorithms. In order to refine the model training process and reduce prediction errors, this work constructs an ensemble learning method that combines the advantages of five regression algorithms. The results show that the integrated model has a smaller prediction error and can be used for subsequent material feature analysis and screening.
Figure 3. The training results of the integrated regression model and the analysis of the prediction accuracy of the model
Point 3: Model analysis - use the SHAP model to suggest screening criteria for organic amine salts
After determining the integrated regression model, this work imported twelve chemical feature descriptors of all passivation materials and thirteen variables of precursor solution concentration of passivation materials into the model for SHAP importance ranking and analysis. The results show that the four material characteristics that have the greatest impact on device efficiency are hydrogen bond donor (Hydrogen bond donor), solution concentration (Concentration), hydrogen atom number (H atom) and lipid-water partition coefficient (MolLogP). Then, according to the characteristic SHAP analysis results, the chemical characteristic screening criteria of passivation materials that help to improve device performance were quantified: Hydrogen bond donor<2, 8 ≤ H atom ≤20, -3.8 ≤ MolLogP ≤ -1.4.
Figure 4. Molecular feature importance analysis and screening criteria for passivation materials
Key point 4: Experimental verification - a substantial breakthrough in device efficiency based on prediction
After determining the screening range of passivation materials, the integrated regression model was used to predict and rank the performance of 112 candidate materials in the PubChem database. Then a variety of materials were selected from the top 10 and bottom 10 passivation materials predicted by the model for experimental verification. Among the materials predicted by the model, FAMACs-based and FAMA-based perovskite solar cells treated with 2-Phenylpropan-1-aminium iodide (2-PPAI) achieved device efficiencies of 22.36% and 24.47%, respectively. The above results show that the model can now screen passivation materials well and provide reference and guidance for experiments. However, the model still has deficiencies, and it is mainly suitable for the qualitative analysis of the trend of passivation materials and the preliminary screening of materials. If quantitative and accurate prediction of efficiency is required, the quantity and accuracy of data still need to be further improved.
Figure 5. Comparison of model prediction and experimental verification results
【Article link】
C.Y. Zhi, S. Wang, S. Sun, C. Li, Z. Li, Z. Wan, H. Wang, * Z. Li, * and Z. Liu, * Machine-Learning-Assisted Screening of Interface Passivation Materials for Perovskite Solar Cells. ACS Energy Letters, 2023, 8, 1424-1433.
https://pubs.acs.org/doi/10.1021/acsenergylett.2c02818
E-mail: info@chemborun.com
Tel: +86-574-87178138 
No. 1558, Jiangnan Road,, Ningbo, Zhejiang, China (Mainland)/31
