Method for predicting mesophase assisted perovskite phase purification by using machine learning and perovskite battery

Abstract

The invention discloses a method for predicting intermediate phase assisted perovskite phase purification by using machine learning and a perovskite cell, and belongs to the technical field of thin-film solar cells. According to the method, a natural language processing (NLP)-based machine learning technology is adopted to analyze documents related to perovskite, a plurality of clusters are obtained, after dimension reduction and visualization processing are carried out on the clusters, cosine value similarity and Euclidean distance of related substances in an intermediate phase and the clusters are calculated on the basis of a visualization processing result, and the similarity and the Euclidean distance of the related substances in the intermediate phase and the clusters are calculated. By means of the method, the appropriate additive for achieving phase purification needed by the perovskite solar cell can be predicted, and compared with the mode that people select the additive in the perovskite from a large number of chemicals, the mode that machine learning is adopted for predicting the additive in the perovskite is faster and more accurate.

Description

Technical Field

[0001] The invention belongs to the technical field of thin-film solar cells and relates to a method for using machine learning to predict an intermediate phase to assist in perovskite phase purification and a perovskite cell. Background Art

[0002] Perovskite solar cells (PSCs) have attracted much attention due to their remarkable optoelectronic properties. Since they were first reported in 2009-2012, the power conversion efficiency (PCE) of PSCs has soared to over 26.5%. This remarkable performance is attributed to their extended carrier diffusion length, prolonged carrier lifetime, tunable band gap, and high light absorption coefficient, making PSCs and their tandem configurations prime candidates for third-generation photovoltaic technology. Among nip structures, the high PCE of formamidinium lead iodide (FAPbI3) is attributed to its ideal band gap of 1.48 eV, which is closely consistent with the optimal band gap (1.33 eV) defined by the Shockley-Queisser limit. FAPbI3 has two different structural phases: the photoactive and black α phase and the non-perovskite yellow δ phase that is stable at room temperature. The α phase and the δ phase are in a competitive state during the crystallization process. Although the δ phase transforms into the α phase at about 130°, it is still challenging to achieve a pure α phase due to the complex phase transition process. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method and a perovskite battery for using machine learning to predict the intermediate phase to assist in perovskite phase purification, so as to solve the problem in the prior art that it is difficult to obtain a pure α phase due to the complex phase change process.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A method for predicting mesophase-assisted perovskite phase purification using machine learning, comprising the following steps: Step 1: Input photovoltaic cell literature into a natural language processing model, and the natural language processing model classifies the same type of related words in the perovskite literature to obtain several clusters; Step 2: Reduce the dimension of the clusters and visualize them; Step 3, based on the visualization processing results, calculating the Euclidean distance between the related words and the keywords in the cluster, the keywords are intermediate or intermediate phase; Step 4: Based on the Euclidean distance, a substance capable of assisting the crystallization and purification of perovskite is obtained.

[0005] A further improvement of the present invention is: Preferably, in step 1, the natural language processing model is word2vec.

[0006] Preferably, in step 1, in the word2vec, the sg module is 1, the word vector dimension is 200, the word window is 10, and the minimum word frequency is 5.

[0007] Preferably, in step 1, before inputting the perovskite literature into the natural language model, the perovskite literature is pre-processed for denoising.

[0008] Preferably, the denoising preprocessing is to remove irrelevant words and punctuation marks in the document and add custom vocabulary.

[0009] Preferably, in step 1, if the frequency of a certain word and a main word appearing simultaneously in the document is greater than a threshold, the certain word is defined as a related word of the main word to obtain related words of the same category.

[0010] Preferably, in step 2, the clusters are reduced in dimension and visualized by t-SNE or UMAP.

[0011] Preferably, in step 3, the similarity between the related words and the keywords is measured by cosine similarity.

[0012] Preferably, in step 4, several substances whose Euclidean distances satisfy the set conditions are obtained, and then the substances that can assist in the crystallization and purification of perovskite are confirmed through experiments.

[0013] A perovskite cell obtained by any of the above-mentioned methods for predicting intermediate phase-assisted perovskite phase purification, wherein the substance capable of assisting perovskite crystallization purification is AG, and the perovskite cell comprises a glass substrate, an electron transport layer, a perovskite film, a hole transport layer and a metal electrode arranged in sequence from bottom to top; The preparation process of the perovskite film is as follows: AG is added to a perovskite precursor solution, and the perovskite precursor solution is dripped onto an electron transport layer to prepare a perovskite film; during the formation of the perovskite film, AGPbI3 crystals are formed as an intermediate phase.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a method for using machine learning to predict intermediate phases to assist in perovskite phase purification. The method uses machine learning technology based on natural language processing (NLP) to quickly analyze literature related to the perovskite, obtain multiple clusters, reduce the dimension of the clusters and visualize them, and then calculate the Euclidean distance between the intermediate phase and related substances in the clusters on the visualization results to obtain the final substance that can promote the crystallization of the perovskite. The method predicts suitable additives for achieving phase purification required for perovskite solar cells. Machine learning predicts additives in perovskites more quickly and accurately than people selecting from a large number of chemicals. The method can quickly explore new intermediate phases to improve the efficiency and operating stability of thin-film solar cells.

[0015] The present invention also discloses a perovskite obtained by a method for assisted perovskite phase purification using machine learning to predict intermediate phases, which uses natural language processing (NLP) to predict that aminoguanidine (AG) hydrochloride has a high correlation with intermediates and is determined to be a promising candidate for forming intermediate phases during nucleation. Experiments have confirmed that AG improves the stability of the precursor solution, can form intermediate phases in the FAPbI3 system to regulate nucleation, and forms a new type of one-dimensional intermediate phase (AGPbI3), suppresses the solvent intermediate phase and promotes the rapid transformation of the α phase, thereby promoting the generation of pure α phase. As a result, the efficiency of the solar cell is increased from 23.99% to 25.46%, which is the highest efficiency reported for TiO2-based single-junction PSCs. Due to the purified α phase, the long-term thermal stability and photostability are significantly enhanced, maintaining 82% of the initial efficiency after aging at 85°C for 1056 hours, and maintaining 86% of the initial efficiency after 1056 hours of illumination in a N2 atmosphere. The method is also able to improve the performance of flexible solar cells, increasing their efficiency from 21.24% to 22.86%. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Embedding and clustering diagrams for perovskite additives, perovskite additives, ETL materials, and hybrid materials; Figure 2 is the relationship diagram between all additive words and the intermediates of the target word; Figure 3 is the cosine similarity graph between related words and intermediates; a larger cosine similarity means that the related words can be clustered as perovskite intermediates; Figure 4 is the Euclidean distance graph between related words and intermediates; short Euclidean distances indicate that words can be clustered as perovskite intermediates; Figure 5 JV curve diagram of flexible device; Figure 6 Comparison diagram, including (a) the binding energy between PbI2 and DMF / FAI / AG respectively; (b) the morphology and XRD diagram of AGPbI3; (c) the lattice structure of AGPbI3; (d) the formation energy of each structure calculated by DFT; (e) GIWAXS of the unannealed control film; (f) GIWAXS of the unannealed film with AG added; (g) the in-situ PL of the control film during annealing; (h) the in-situ PL of the target film during annealing.

[0017] Figure 7 The figures are comparison diagrams, where (a) shows the JV curve of the solar cell under reverse scanning; (b) shows the statistical box plot of PCE; and (c) shows the steady-state output when measuring the photocurrent density and PCE. Figure 8 Calculation diagram of Euclidean distance and cosine similarity; (a) weighted average of cosine similarities of “better”, “additive”, “FAPbI3” and “nucleation”; (b) cosine similarity of “better”; (c) cosine similarity of additive; (d) cosine similarity of FAPbI3; (e) cosine similarity of “nucleation”; (f) Euclidean distance of “better”; (f) Euclidean distance of “better”; (g) Euclidean distance of additive; (h) Euclidean distance of FAPbI3; (i) Euclidean distance of “nucleation”. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below in conjunction with the accompanying drawings: In order to enable those skilled in the art to understand the characteristics and effects of the present invention, the following is a general description and definition of the terms and expressions mentioned in the specification and claims. Unless otherwise specified, all technical and scientific terms used in the text are the common meanings understood by those skilled in the art for the present invention. In the event of a conflict, the definition in this specification shall prevail.

[0019] In this document, unless otherwise specified, “includes,” “including,” “contains,” “has,” or similar terms cover the meanings of “consisting of” and “mainly consisting of,” for example, “A includes a” covers the meanings of “A includes a and other” and “A only includes a.”

[0020] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0021] The following examples use conventional instruments and equipment in the art. The experimental methods in the following examples where specific conditions are not specified are usually carried out under conventional conditions or under conditions recommended by the manufacturer. The various raw materials used in the following examples are conventional commercial products unless otherwise specified, and their specifications are conventional specifications in the art. In the specification of the present invention and the following examples, unless otherwise specified, "%" means weight percentage, "part" means weight part, and ratio means weight ratio.

[0022] The present invention discloses a method for using machine learning to predict intermediate phases to assist in perovskite phase purification, the method comprising the following steps: Step 1: Input the photovoltaic device literature into the natural language processing model (NLP). The natural language processing model classifies the same type of related words in the perovskite literature to obtain several clusters. Step 2: Reduce the dimension of the clusters and visualize them; Step 3, based on the visualization processing results, calculating the Euclidean distance between the related words and the keywords in the cluster, the keywords are intermediate or intermediate phase; Step 4: Based on the Euclidean distance, a substance capable of assisting the crystallization and purification of perovskite is obtained.

[0023] In some embodiments of the present invention, in step 1, the natural language processing model is a word2vec model, which is a model for generating word embeddings. The Word2Vec model maps each word to a vector in a high-dimensional space by learning the contextual relationship between words, and these vectors can capture the semantic information and contextual information of the word. The word2vec model outputs a high-dimensional vector to obtain a number of clusters in the high-dimensional vector.

[0024] In some implementation schemes of this aspect, in step 1, literature abstracts are downloaded, and an API (application programming interface) is used to obtain a database of 1.18 million literature abstracts in physics, chemistry, and materials; training is performed directly through the abstracts in the literature, and the amount of data input to the model is reduced when the key points of each document can be obtained.

[0025] In some embodiments of the present invention, in step 1, the literature data is subjected to denoising preprocessing, and the denoising preprocessing specifically includes deleting stop words and punctuation marks, adding specific vocabulary to the preprocessed training set, and then converting it into a database format suitable for word embedding training; the database in step 1 is deleted of irrelevant words (stop words) and irrelevant punctuation marks (for example, .!, ; and other symbols unrelated to chemical formulas), and custom vocabulary (mainly "solar cell") is added, and then converted into a database format suitable for word embedding training, including stop words, keywords, and custom dictionaries.

[0026] In some embodiments of the present invention, in step 1, the preprocessed training set is input into the natural language processing model, and the parameters are adjusted. The word embedding module is Word2vec, the sg module is set to 1, the word vector dimension is 200, and each word is converted into a 200-dimensional word vector; the word window windows is set to 10, and the minimum word frequency is set to 5. As long as the low-frequency words with a number of occurrences less than 5 are not involved in the word vector embedding.

[0027] Specifically, in step 1, clusters are obtained that are mostly related words of a certain type. For example, with electron transport layer as the main word, the above model can screen out related words with high frequency of co-occurrence with electron transport layer from multiple documents, and then regard the related words as the same category; similarly, hole transport layer, perovskite film (perovskite absorption layer, perovskite layer), and additives can be screened out as main words, and the related words corresponding to each main word can be screened out as the same category.

[0028] In some embodiments of the present invention, in step 2, dimensionality reduction and visualization processing are performed using relevant algorithms such as t-SNE, UMAP, etc., so that high-dimensional clusters can be reflected on a plane.

[0029] In some implementation schemes of the present invention, in step 3, the cosine value and Euclidean distance between related words and keywords in the same type of clusters are calculated based on the visualized results.

[0030] Exemplarily, the language model is built based on millions of scientific papers in the fields of chemistry and materials. The databases mainly come from Web of Science and SpringerLink, which limit searches to specific fields such as "materials" or "perovskite solar cells" and contain abstracts of about one million articles from 1960 to 2024. The Natural Language Toolkit was used for text preprocessing (see the supporting information for details). In step 2, t-distributed random neighbor embedding (t-SNE) is used to reduce the high-dimensional word vector (200 dimensions) to two dimensions. Word vectors with high similarity will be closer on the two-dimensional plane, while word vectors with low similarity will be farther apart, such as Figure 1 Clustering on a two-dimensional plane reveals that the materials of TiO2 and SnO2 are clustered together, clearly indicating their role in the electron transport layer materials. FAPbI3, MAPbI3, and CsPbBr3 represent the absorption layer, while the words “chloride,” “acetate,” and “bromide” are associated with “additives” (see Figure 2 ), and therefore clustered as a perovskite additive. This demonstrates the accuracy of the model in clustering the materials.

[0031] In step 3, the following two metrics were used to cluster the materials: 1) Cosine similarity: This metric measures the cosine of the angle between two word vectors (material formula and target application) in the embedding space, which is used to evaluate the similarity between word embeddings. Higher cosine similarities indicate that the words are more semantically similar. 2) Quantifying clustering quality through Euclidean distance: The quality of clustering is evaluated by examining the compactness of the formed clusters. This involves analyzing the correlation between items within each cluster.

[0032] The cosine similarity between the terms additive and FAPbI3, mesophase, nucleation, etc. is calculated. Figure 3 In the data, some additives such as caffeine have weak correlations with mesophase and nucleation, indicating that these additives may not form mesophase during nucleation, but AG has the highest similarity with “mesophase”. In addition, the Euclidean distance is used to measure the distance between embeddings of different words or clusters, see Figure 4 , which helps to understand the spatial relationship between word vectors in the embedding space. Calculating the Euclidean distance between these materials and "intermediate phase", it is found that the Euclidean distance between AG and "intermediate phase" is the shortest. AG has the shortest cosine similarity (see Figure 3 ) and clustering (see Figure 4 ). In addition, the cosine similarity and Euclidean distance of these materials to the “better” ones were calculated to ensure that they were positively correlated (see Figure 8 ). All the above calculation results indicate that AG may act as an additive in the formation of FAPbI3 intermediate phases. Therefore, AG was chosen to explore new intermediate phases for crystallization optimization in FAPbI3 perovskite solar cells.

[0033] In some embodiments of the present invention, in step 4, based on the calculations in the above steps, several additives that meet the conditions can be obtained, and subsequently further experimental verification can be performed to determine which substance or substances are the optimal additives.

[0034] Furthermore, the preparation of a perovskite solar cell using the additive obtained in the above steps comprises the following steps: S1, fluorine-doped tin oxide (FTO) glass was selected as the substrate. FTO was ultrasonically cleaned in a cleaning concentrate, deionized water and ethanol in sequence, the cleaned substrate was dried with a nitrogen stream, and treated in an ultraviolet ozone cleaner for 15 minutes, after which an electron transport layer was deposited on the substrate; specifically, the material of the electron transport layer in S1 was any one of TiO2, ZnO and SnO2.

[0035] Exemplarily, the specific process of preparing the TiO2 electron transport layer is: placing the substrate in a 0.2M TiCl4 aqueous solution at 70°C for 60 minutes, and then washing it three times alternately with deionized water and ethanol; annealing the substrate at 200°C in air for 30 minutes.

[0036] S2. Dissolve AG in a mixed solution of N,N-dimethylformamide and dimethyl sulfoxide (mixing volume ratio 4:1), and stir to dissolve to obtain a liquid crystal additive solution with a concentration of 6 mg / mL; according to the molar amounts of PbI2, NH2CH=NH2I and CH3NH3Cl, which are 1.20 ~1.80 mmol, 1.20 ~1.80 mmol and 0.42 ~0.63 mmol respectively, they are mixed and dissolved in 1 mL of the additive solution to obtain a perovskite precursor solution containing the additive.

[0037] The structural formula of the additive material CH3NH3Cl is shown in formula (1): (1) S3, dropping about 80 μL of perovskite precursor on the TiO2 / FTO / glass substrate, and using the reverse solution method with ether as the anti-solvent to form a perovskite film; About 80 μL of perovskite precursor was dropped on the TiO2 / FTO / glass substrate, and the perovskite film was formed by the anti-solution method with ether as the anti-solvent. The solution was spun at 3000 rpm for 10 seconds and then at 5000 rpm for 30 seconds. 20 seconds before the end of 5000 rpm, 600 μL of ether was quickly dropped on the surface. Next, the film was annealed at 150 ° C for 15 minutes. During the formation of the perovskite film, the intermediate phase AGPbI3 is generated. As the annealing proceeds, the intermediate phase AGPbI3 disappears. The final perovskite film is composed of FAPbI3 and MACl, with AG at the grain boundary.

[0038] Preferably, before spin coating, S3 is treated in ultraviolet ozone for 20 minutes to improve wettability.

[0039] S4, generating a hole transport layer made of Spiro-OMeTAD, poly[bis(4-phenyl)(2,4,6-trimethylphenyl)amine] or nickel oxide on the perovskite absorption layer on the glass substrate; In S4, a hole transport layer made of Spiro-OMeTAD is generated, which specifically includes the following steps: Spin coat 50 μL of Spiro-OMeTAD solution at 5000 rpm for 30 s to form a hole transport layer.

[0040] S5, generating an electrode by evaporation on the hole transport layer of the glass substrate, and finally obtaining a perovskite solar cell; Preferably, the electrode is a gold electrode with a thickness of 80 nm and an effective area of ​​0.09 cm 2 .

[0041] Engineering of mesophases is an effective method for regulating nucleation and optimizing crystallization. However, most methods for selecting mesophase-forming additives are still in their infancy and rely on trial and error. Selecting suitable molecules from the millions of candidate molecules in the literature is challenging and usually requires researchers to spend a lot of time and effort reading. The present invention uses language machine learning (such as machine learning based on natural language processing) to interpret the literature, so that suitable additives can be screened out from a large amount of text. In a specific embodiment of the present invention, for the above reasons, machine learning technology based on natural language processing (NLP) is used to quickly analyze about 10 in a short time (about three hours). 6 Abstracts related to chemistry and materials science are presented, and additives used for mesophase formation in this study are predicted.

[0042] The present invention uses natural language processing (NLP) to predict that AG can form an intermediate phase in the FAPbI3 system to regulate nucleation. AG forms a new one-dimensional intermediate phase during the nucleation process, which suppresses the solvent intermediate phase and promotes the rapid transformation of the α phase, suppresses the formation of the δ phase, and thus promotes the generation of a pure α phase. Through single crystal analysis, in situ photoluminescence (PL) characterization, and density functional theory (DFT) calculations, it is found that the formation of the new intermediate phase enhances solvent volatilization during the annealing stage. The enhancement of solvent volatilization reduces solvent residues, thereby producing a high-crystallinity film and releasing residual stress during the annealing process.

[0043] In order to allow those skilled in the art to further understand the technical solution of the present invention, the technical solution of the present invention is further described in detail below through specific embodiments and comparative examples.

[0044] Example 1 The method of the present invention for preparing a perovskite solar cell by using the method based on machine learning to predict the additive AG to form an intermediate phase to assist the perovskite phase purification is as follows: Step 1, first, fluorine-doped tin oxide (FTO) glass was selected as the substrate. FTO was ultrasonically cleaned in a cleaning concentrate, deionized water, and ethanol for 30 minutes in sequence. Then, the cleaned substrate was dried with a nitrogen stream and treated in a UV ozone cleaner for 15 minutes; then, the substrate was placed in a 0.2M TiCl4 aqueous solution at 70°C for 60 minutes, and then alternately cleaned with deionized water and ethanol three times; finally, the substrate was annealed at 200°C in air for 30 minutes; Step 2, dissolving AG in a mixed solution of N,N-dimethylformamide and dimethyl sulfoxide, stirring and dissolving to obtain a liquid crystal additive solution with a concentration of 6 mg / mL; mixing and dissolving PbI2, NH2CH=NH2I and CH3NH3Cl in molar amounts of 1.20 mmol, 1.20 mmol and 0.42 mmol respectively in 1 mL of the additive solution to obtain a perovskite precursor solution containing the additive; Step 3, about 80 μL of perovskite precursor was dropped on the TiO2 / FTO / glass substrate, and the perovskite film was formed by the anti-solution method with ether as the anti-solvent. The solution was spun at 3000 rpm for 10 seconds and then at 5000 rpm for 30 seconds. 20 seconds before the end of 5000 rpm, 600 μL of ether was quickly dropped on the surface. Next, the film was annealed at 150°C for 15 minutes.

[0045] Step 4: Spin coat 50 μL of Spiro-OMeTAD solution at 5000 rpm for 30 seconds to form a hole transport layer.

[0046] Step 5, generating an electrode by evaporation on the hole transport layer of the glass substrate, preferably a gold electrode, the thickness of the gold electrode being 80 nm, and finally obtaining a perovskite solar cell; Comparative Example 1 The method of AG changing the crystallization kinetics was not used to prepare perovskite solar cells.

[0047] Except for step 2, the other steps are the same as those in Example 1, except that AG is not added in step 2 of comparative example 1, and a perovskite precursor solution without AG is prepared. Step 2 is as follows: PbI2, NH2CH=NH2I (FAI) and CH3NH3Cl (MACl) are mixed and dissolved in 1 mL of a mixed solvent (DMF and DMSO volume ratio 4:1) in molar amounts of 1.20 mmol, 1.20 mmol and 0.42 mmol, respectively, to prepare a perovskite precursor solution.

[0048] The products of the perovskite absorption layer formed in step 2 of the above-mentioned embodiments and comparative examples are named as perovskite films. The performance of the perovskite films prepared in Example 1 and Comparative Example 1 is characterized and analyzed from three aspects: introduction of machine learning prediction, phase purification and intermediate phase generation, and performance and stability improvement.

[0049] The language model is built on millions of scientific papers in the field of materials. The database is provided by SpringerLink, which restricts searches to specific fields, such as "materials," and contains about 10,000 papers between 1960 and 2020. 6Abstract of the article. The Natural Language Toolkit was used for text preprocessing. The high-dimensional word vector (200 dimensions) was reduced to two dimensions using t-distributed stochastic neighbor embedding (t-SNE). Word vectors with high similarity are closer in the two-dimensional plane, while word vectors with low similarity are farther apart ( Figure 1 ). Clustering on the two-dimensional plane shows that materials such as TiO2 and SnO2 are clustered into one category, which clearly indicates their role as electron transport layer materials. FAPbI3, MAPbI3, and CsPbBr3 represent the absorption layer, while previously reported additives such as chloride, caffeine, and thiourea are clustered as perovskite additives. This shows that the model is accurate in clustering materials. In addition, the presence of AG in the additive category indicates its potential as a candidate for perovskite additives. Next, taking the words in the graph (FAPbI3 / intermediate / nucleation…) as the target, the cosine similarity score of all words in the model was calculated ( Figure 2 ). The color intensity represents the cosine value, with the darkest purple representing a value of 1.0, indicating a perfect correlation between the two words, and 0 indicating no correlation. The cosine similarity between AG and the terms FAPbI3, intermediate, and nucleation was calculated. Figure 3 In the figure, AG appears in the darker region, indicating a high similarity between AG and “intermediate”. In addition, AG shows a high correlation with nucleation and FAPbI3, indicating that AG may act as an additive in the formation of the FAPbI3 intermediate phase ( Figure 3 ).

[0050] (2) Phase purification and intermediate phase generation According to the analysis of the NLP model, AG may form a new intermediate phase. Using density functional theory (DFT) calculations and structural analysis of single crystals, the formation of the intermediate phase was confirmed. Figure 6 As shown in Figure (a), DFT calculations further confirm that AG can form a new intermediate phase and inhibit the intermediate phase formation of PbI2·DMF. The binding energy of AG with PbI2 (-1.37 eV) is lower than that of PbI2 and DMF (-0.74 eV), indicating that PbI2 is more likely to react with AG to form a new intermediate phase. This new intermediate phase can inhibit the formation of the FAI·PbI2·DMF intermediate phase, which can be quickly transformed into the δ phase. Therefore, AG reduces the interaction between PbI2 and DMF, thereby preventing the formation of the δ phase. AG and PbI2 jointly generate the crystal structure of the new intermediate phase, from Figure 6 As can be seen from Figure (b) in the figure, AGPbI3 is a yellow crystal. The crystal structure of AGPbI3 is as follows: Figure 6As shown in Figure (c), the structure is similar to the δ phase, and both structures are composed of coplanar octahedrons. The spacing between the (110) planes in AGPbI3 is 4.48A. Figure 6 As shown in Figure (d), the DFT calculation results show that the formation energy of AGPbI3 (-1.72 eV) is lower than that of the α phase (-1.48 eV), indicating that AGPbI3 has a thermodynamic advantage as an intermediate phase of FAPbI3 during annealing. During annealing, the process of AG·PbI2 transforming into the intermediate phase AGPbI3 in the precursor solution can compete with the δ phase. Based on this, the design principles of molecules for intermediate phase formation are: (1) From large-scale text analysis of machine learning, molecules that are positively correlated with intermediate phase formation can be quickly explored; (2) The binding energy between the molecule and PbI2 is stronger than the binding energy between the solvent and PbI2, which helps to overcome the solvent intermediate phase; (3) The formation energy of the intermediate phase should be lower than that of the final solid, which is more beneficial from a thermodynamic point of view.

[0051] Figure 5 This is the performance of the flexible solar cell obtained by this method. As can be seen from the figure, this method can also improve the performance of flexible solar cells, increasing their efficiency from 21.24% to 22.86%.

[0052] To explore the effect of AG on crystallization, the following characterizations were performed. Grazing-incidence wide-angle X-ray scattering (GIWAXS, Figure 6 Figures (e) and (f) in Figure 3 characterize the effect of AG on the wet film before annealing. The interplanar spacing d is 2π / q, and d=6.3 Å at q=1.00 Å−1, corresponding to the (001) plane of the α phase. The control wet film mainly presents the δ phase (q=0.83 Å−1) with many impurity peaks, while the wet film containing AG significantly reduces the formation of the δ phase and presents a strong peak at q=1.00 Å−1, corresponding to the α phase. This indicates that the addition of AG promotes the formation of the α phase. Figure 6 (g) and (h) show the in situ measurement results during the film annealing process.

[0053] (3) Improved performance and stability Finally, single-junction nip PSCs were fabricated using the FTO / TiO2 / FAPbI3 / Spiro-OMeTAD / Au structure, and their photovoltaic performance was evaluated. The photocurrent-voltage (JV) curves of the optimized devices are shown in Figure 2. Figure 7As shown in (a) in the figure, the detailed photovoltaic parameters are shown in Table 1. The PCE of Comparative Example 1 is 23.99%, the open circuit voltage is 1.160 V, and the fill factor is 80.61%. The PCE of Example 1 is improved to 25.46%, the open circuit voltage is 1.184 V, and the fill factor is 83.03%, which is the highest among TiO2-based single-junction PSCs. The photovoltaic performance of 20 independent devices was independently evaluated, as shown in Table 1. Figure 7 As shown in (b), the average PCE increases from 23.5% to 25.2%. Figure 7 (c) shows the IT curves of PSCs under continuous illumination at the maximum power point, with negligible changes after 600 s, indicating good stability.

[0054] Table 1. Performance test results of solar cells prepared in Example 1 and Comparative Example 1

[0055] In summary, the method proposed in the present invention, which predicts additive AG based on machine learning to form an intermediate phase to assist perovskite phase purification and improve the performance and stability of perovskite solar cells, not only improves the device efficiency of the perovskite battery, but also improves the stability of the device.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for using machine learning to predict mesophase-assisted perovskite phase purification, characterized in that: The following steps are involved: Step 1: Input photovoltaic cell literature into a natural language processing model, and the natural language processing model classifies the same type of related words in the perovskite literature to obtain several clusters; Step 2: Reduce the dimension of the clusters and visualize them; Step 3, based on the visualization processing results, calculating the Euclidean distance between the related words and the keywords in the cluster, the keywords are intermediate or intermediate phase; Step 4: Based on the Euclidean distance, a substance capable of assisting the crystallization and purification of perovskite is obtained.

2. The method of using machine learning to predict mesophase-assisted perovskite phase purification according to claim 1, characterized in that: In step 1, the natural language processing model is word2vec.

3. The method of using machine learning to predict mesophase-assisted perovskite phase purification according to claim 2, characterized in that: In step 1, in the word2vec, the sg module is 1, the word vector dimension is 200, the word window is 10, and the minimum word frequency is 5.

4. The method of claim 1, wherein the method comprises: In step 1, the perovskite literature is pre-processed for denoising before being input into the natural language model.

5. The method of using machine learning to predict mesophase-assisted perovskite phase purification according to claim 4, characterized in that: The denoising preprocessing is to remove irrelevant words and punctuation marks in the document and add custom vocabulary.

6. The method of claim 1, wherein the method comprises: In step 1, if the frequency of a certain word and a main word appearing simultaneously in the document is greater than a threshold, the certain word is defined as a related word of the main word to obtain related words of the same category.

7. The method of claim 1, wherein the method comprises: In step 2, the clusters are reduced in dimension and visualized by t-SNE or UMAP.

8. The method of claim 1, wherein the method comprises: In step 3, the similarity between the related words and the keywords is measured by cosine similarity.

9. The method of claim 1, wherein the method comprises: In step 4, several substances whose Euclidean distances satisfy the set conditions are obtained, and then the substances that can assist in the crystallization and purification of perovskite are confirmed through experiments.

10. A perovskite cell obtained by the method for predicting mesophase-assisted perovskite phase purification according to any one of claims 1 to 9, characterized in that: The substance capable of assisting the crystallization and purification of perovskite is AG, and the perovskite cell comprises a glass substrate, an electron transport layer, a perovskite film, a hole transport layer and a metal electrode arranged in sequence from bottom to top; The preparation process of the perovskite film is as follows: AG is added to a perovskite precursor solution, and the perovskite precursor solution is dripped onto an electron transport layer to prepare a perovskite film; during the formation of the perovskite film, AGPbI3 crystals are formed as an intermediate phase.

Images