Design method and system for light stabilizer of perovskite photovoltaic device

Abstract

The invention discloses a design method and system for a light stabilizer of a perovskite photovoltaic device, and relates to the technical field of new energy photovoltaic devices, and the method comprises the steps: executing a stabilizer candidate generation task through a generative language model which is finely adjusted through a QO2MOL data set and a dipole design scheme, obtaining a preset number of stabilizer molecule candidates, and carrying out the calculation of the stabilizer molecule candidates; according to a pre-constructed prediction model, constructing a molecule vector corresponding to a preset number of stabilizer molecule candidates, according to the molecule vector and a sorting algorithm, obtaining an expected number of selected candidates, and marking the expected number of selected candidates through a chemical informatics tool, and selecting a target molecule meeting a dipole intensity threshold value from the marked selected candidates with the expected number, taking the target molecule as a target stabilizer configuration, and designing the perovskite photovoltaic device light stabilizer through the target stabilizer configuration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy photovoltaic devices, more particularly to a design method and system of a perovskite photovoltaic device light stabilizer. BACKGROUND

[0002] The energy conversion efficiency, i.e., the certified efficiency, of a perovskite photovoltaic device refers to the ability of a perovskite solar cell to convert solar energy into electrical energy. The energy conversion efficiency (PCE) is an important indicator for measuring the performance of a solar cell, representing the ratio of the electrical energy output by the cell to the energy of incident sunlight.

[0003] Although the energy conversion efficiency of a perovskite photovoltaic device has reached 22.8%, the stability of the microstructure of the device is still a key factor limiting the working efficiency and lifespan of the photovoltaic device when the device is continuously working under light. Factors affecting the stability of the microstructure of the device include halogen iodine ion interface defects, etc.

[0004] Therefore, how to improve the stability of the microstructure of the device is a problem that needs to be solved by the present application. SUMMARY

[0005] Therefore, the present application discloses a design method and system of a perovskite photovoltaic device light stabilizer, aiming to design a perovskite photovoltaic device light stabilizer by using a molecule with the highest dipole strength as a target stabilizer configuration, so as to improve the stability of the microstructure of the device.

[0006] In order to achieve the above-mentioned purpose, the disclosed technical solution is as follows:

[0007] The first aspect of the present application discloses a design method of a perovskite photovoltaic device light stabilizer, which comprises:

[0008] A generated language model is fine-tuned through a QO2MOL dataset and a dipole design scheme, a stabilizer candidate generation task is performed, and a preset number of stabilizer molecule candidates are obtained;

[0009] A molecular vector corresponding to the preset number of stabilizer molecule candidates is constructed according to a pre-constructed prediction model, and a selected number of selected candidates is obtained according to the molecular vector and a sorting algorithm;

[0010] The selected number of selected candidates is labeled by a chemical informatics tool;

[0011] A target molecule satisfying a dipole strength threshold is selected from the labeled selected number of selected candidates;

[0012] The target molecule is configured as a target stabilizer configuration, and a perovskite photovoltaic device light stabilizer is designed through the target stabilizer configuration.

[0013] Preferably, the fine-tuning process of the generative language model comprises:

[0014] In the pre-training stage, a corpus for fine-tuning of the generative language model is obtained from the QO2MOL dataset; wherein the QO2MOL dataset comprises at least SMILES expressions and molecular property labels; the molecular property labels are labels associated with the dipole strength scheme; the molecular property labels comprise at least molecular dipole strength;

[0015] The tokenizer is processed into an input form of the generative language model; wherein the tokenizer comprises a dictionary sequence and a mapping rule; the mapping rule is a mapping rule for converting SMILES expressions in the QO2MOL dataset into molecular vectors;

[0016] The corpus for fine-tuning of the generative language model and the input form are imported into the fine-tuning process of the generative language model.

[0017] Preferably, the fine-tuned generative language model through the QO2MOL dataset and the dipole design scheme performs a stabilizer candidate generation task to obtain a preset number of stabilizer molecule candidates, comprising:

[0018] The fine-tuned generative language model through the QO2MOL dataset and the dipole design scheme performs a stabilizer candidate generation task;

[0019] In the process of performing the stabilizer candidate generation task, a preset number of stabilizer molecule candidates are randomly generated.

[0020] Preferably, the process of constructing the prediction model comprises:

[0021] An extended dataset of the QO2MOL dataset is obtained;

[0022] According to the open source machine learning tool, a data loading method is selected;

[0023] According to the data loading method, the QO2MOL dataset and the extended dataset are loaded to extract valid molecular structure formulas from the QO2MOL dataset and molecular dipole strength labels from the extended dataset;

[0024] According to a preset ratio, the QO2MOL dataset and the extended dataset are respectively divided to obtain a training set, a test set and a validation set;

[0025] According to the molecular vector construction method and the valid molecular structure formula, a molecular vector is obtained;

[0026] inputting the molecular vector into the initial prediction model to obtain an output result of the initial prediction model;

[0027] obtaining a loss function by the molecular dipole strength label and the output result of the initial prediction model within any time period;

[0028] training the initial prediction model by the loss function and the training set;

[0029] testing the trained initial prediction model by the test set and verifying the tested initial prediction model by the verification set to complete the process of constructing the prediction model.

[0030] Preferably, the molecular vector corresponding to the preset number of stabilizer molecule candidates is constructed according to the pre-constructed prediction model, and the expected number of selected candidates is obtained according to the molecular vector and the sorting algorithm, including:

[0031] inputting the preset number of stabilizer molecule candidates into the pre-constructed prediction model to obtain a prediction model output result, and taking the prediction model output result as a property label of the candidate;

[0032] constructing a molecular vector corresponding to the preset number of stabilizer molecule candidates according to the property label of the candidate and a molecular annotation tool;

[0033] calculating the distance similarity between the molecular vectors corresponding to the preset number of stabilizer molecule candidates;

[0034] constructing an adjacency matrix on the distance similarity by a recommendation algorithm, and constructing a graph network based on the adjacency matrix;

[0035] obtaining the expected number of selected candidates by a sorting algorithm realized based on the graph network and the molecular vector.

[0036] Preferably, the expected number of target molecules is selected from the labeled expected number of selected candidates, including:

[0037] selecting the expected number of target molecules from the labeled expected number of selected candidates by a density functional theory framework.

[0038] The second aspect of the present application discloses a design system of a perovskite photovoltaic device light stabilizer, the system comprising:

[0039] a first acquisition unit configured to execute a stabilizer candidate generation task by using a generated language model fine-tuned by a QO2MOL dataset and a dipole design scheme, and obtain a preset number of stabilizer molecule candidates.

[0040] a second obtaining unit, configured to construct a molecular vector corresponding to the preset number of stabilizer molecule candidates according to a prediction model constructed in advance, and obtain a selected number of selected candidates according to the molecular vector and a sorting algorithm;

[0041] a labeling unit, configured to label the selected number of selected candidates by using a cheminformatics tool;

[0042] a selecting unit, configured to select a target molecule satisfying a dipole strength threshold from the labeled selected number of selected candidates;

[0043] a designing unit, configured to use the target molecule as a target stabilizer configuration, and design a perovskite photovoltaic device light stabilizer by using the target stabilizer configuration.

[0044] Preferably, the first obtaining unit of the fine-tuning process of the generative language model comprises:

[0045] a first obtaining module, configured to obtain corpus for fine-tuning of the generative language model from the QO2MOL dataset in a pre-training stage; wherein the QO2MOL dataset at least comprises SMILES expressions and molecular property labels; the molecular property labels are labels associated with a dipole strength scheme; and the molecular property labels at least comprise molecular dipole strength;

[0046] a generating module, configured to process a tokenizer into an input form of the generative language model; wherein the tokenizer comprises a dictionary sequence and a mapping rule; and the mapping rule is a mapping rule for converting the SMILES expressions in the QO2MOL dataset into molecular vectors;

[0047] an importing module, configured to import the corpus for fine-tuning of the generative language model and the input form into the fine-tuning process of the generative language model.

[0048] Preferably, the first obtaining unit comprises:

[0049] an executing module, configured to perform a stabilizer candidate generation task by using the fine-tuned generative language model, the QO2MOL dataset, and the dipole design scheme;

[0050] a random generating module, configured to randomly generate a preset number of stabilizer molecule candidates in the process of performing the stabilizer candidate generation task.

[0051] Preferably, the second obtaining unit of the process of constructing the prediction model comprises:

[0052] a second obtaining module, configured to obtain an extended dataset of the QO2MOL dataset;

[0053] The selection module is configured to select a data loading mode according to an open-source machine learning tool;

[0054] The loading extraction module is configured to load the QO2MOL dataset and the extended dataset according to the data loading mode, to extract effective molecular structural formulas from the QO2MOL dataset, and to extract molecular dipole strength labels from the extended dataset;

[0055] The division module is configured to divide the QO2MOL dataset and the extended dataset into a training set, a test set, and a validation set according to a preset ratio;

[0056] The third acquisition module is configured to acquire a molecular vector according to a molecular vector construction mode and the effective molecular structural formula;

[0057] The fourth acquisition module is configured to input the molecular vector into an initial prediction model to obtain an output result of the initial prediction model;

[0058] The fifth acquisition module is configured to acquire a loss function by using the molecular dipole strength label and the output result of the initial prediction model within any time period;

[0059] The training module is configured to train the initial prediction model by using the loss function and the training set;

[0060] The test and verification module is configured to test the trained initial prediction model by using the test set, and to verify the tested initial prediction model by using the validation set, so as to complete the process of constructing a prediction model.

[0061] According to the above technical solution, the application discloses a design method and system for a perovskite photovoltaic device light stabilizer. The QO2MOL dataset and the dipole design scheme are used to fine-tune the generated language model, perform the stabilizer candidate generation task, obtain a preset number of stabilizer molecular candidates, construct molecular vectors corresponding to the preset number of stabilizer molecular candidates according to the pre-constructed prediction model, and obtain a desired number of selected candidates according to the molecular vectors and the sorting algorithm. The selected candidates are marked by using a chemical informatics tool, and a target molecule satisfying a dipole strength threshold is selected from the marked selected candidates. The target molecule is used as a target stabilizer configuration, and the perovskite photovoltaic device light stabilizer is designed by using the target stabilizer configuration.

[0062] The beneficial effects of the present application: the design scheme of the double-driven perovskite photovoltaic device light stabilizer generated by the self-developed QO2MOL molecular scientific corpus and the generated language model, through the design scheme of the double-driven perovskite photovoltaic device light stabilizer, the polarity properties of the molecule are described by combining the molecular dipole strength, such as dipole rate, polarization intensity, etc. The candidate molecules of the stabilizer are designed, and the highest dipole strength molecule is finally selected through the sorting algorithm, prediction model screening and calculation chemistry tool fine screening. The highest dipole strength molecule selected shows effective passivation and stability improvement effect on the surface iodine defects of the perovskite device. The highest dipole strength molecule is used as the target stabilizer configuration, and the perovskite photovoltaic device light stabilizer is designed through the target stabilizer configuration to improve the stability of the device microstructure. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0064] Figure 1 The flowchart of the design method of the perovskite photovoltaic device light stabilizer disclosed by the embodiments of the present application is shown.

[0065] Figure 2 The schematic diagram of the large-scale language model architecture disclosed by the embodiments of the present application is shown.

[0066] Figure 3 The schematic diagram of the input adjacency matrix of the recommendation algorithm disclosed by the embodiments of the present application is shown.

[0067] Figure 4 The simulation and verification of the physical and chemical processes of the light stabilizer in the perovskite device disclosed by the embodiments of the present application are shown.

[0068] Figure 5 The structural schematic diagram of the design system of the perovskite photovoltaic device light stabilizer disclosed by the embodiments of the present application is shown.

[0069] Figure 6 The structural schematic diagram of the electronic device disclosed by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0070] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0071] In the present application, the term "comprising", "containing" or any other variant thereof is intended to cover the non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0072] From the above background, although the energy conversion efficiency of the perovskite photovoltaic device has reached 22.8%, when the perovskite photovoltaic device works continuously under light, the stability of the microstructure of the device is still a key factor limiting the working efficiency and service life of the photovoltaic device. The factors affecting the stability of the microstructure of the device include halogen iodine ion interface defects, etc. Therefore, how to improve the stability of the microstructure of the device is a problem to be solved by the present application.

[0073] In order to solve the above problems, the present application discloses a design method and system of a perovskite photovoltaic device light stabilizer, which adopts a self-developed QO2MOL molecular scientific corpus and a double-driven perovskite photovoltaic device light stabilizer design scheme for generating a real language model. Through the double-driven perovskite photovoltaic device light stabilizer design scheme, the polarity properties of the molecule are described by combining the molecular dipole strength, such as dipole rate and polarization strength, to design stabilizer molecule candidates. Through sorting algorithm, prediction model screening and calculation chemistry tool fine screening, the highest dipole strength molecule is finally selected. The highest dipole strength molecule selected shows effective passivation and stability improvement effect on the surface iodine defects of the perovskite device. The highest dipole strength molecule is used as a target stabilizer configuration, and the perovskite photovoltaic device light stabilizer is designed through the target stabilizer configuration to improve the stability of the microstructure of the device. The specific implementation manner is specifically described in the following embodiments.

[0074] Reference Figure 1 As shown in the figure, the design method of the perovskite photovoltaic device light stabilizer disclosed in the embodiments of the present application mainly includes the following steps:

[0075] S101: Perform the stabilizer candidate generation task by the fine-tuned generative language model based on the QO2MOL dataset and the dipole design scheme, to obtain a preset number of stabilizer molecule candidates.

[0076] It should be noted that the QO2MOL dataset is a large-scale first-principle molecular property dataset for material science developed by the applicant, which is constructed through 400 million core hours of high-precision calculation, and has the characteristics of high precision, full element type, and large scale.

[0077] The generative language model is a language model with generation inference capability. The generative language model includes attention mechanism, multi-head attention mechanism, Mask operation, etc. The architecture of the generative language model includes but is not limited to decode-only architecture, encode-decode architecture, etc.

[0078] The generative language model takes the output of the last hidden layer as the model output and performs Tokenzier decoding. Sample generation includes numerical parameters, no need for prompt input, parameter selection, etc.

[0079] The generative language model of the present scheme, i.e., the large-scale language model, has an architecture including but not limited to the self-developed material science large model molecular design module, the open-source hybrid inference model (Qwen3MoE), and the open-source artificial intelligence model Gemma. The specific large-scale language model architecture is shown in Figure 2 .

[0080] Figure 2 In the present scheme, the large-scale language model architecture is a large model architecture that meets the scaling rate. The generative language model protects the QO2Mol application scenario / dipole. The generative language model and the Tokenzier data processing tool process the data into an input vector of the transformer open-source language model.

[0081] Through the standard workflow and common structure information of the large-scale language model, the value in this scenario of different architectures is covered.

[0082] The workflow of the large-scale language model is as follows:

[0083] 1. Tokenzier data preprocessing, such as vectorizing the molecule SMILES;

[0084] 2. Input to the embedding layer of the large-scale language model;

[0085] 3. Output according to the last_hidden_layerji result;

[0086] 4. Decode the molecule vector into SMILES (including random output) using tokenizer.

[0087] 5. Large-scale language model training based on the last_hidden_layerji result and autogress loss of the above 3;

[0088] The generative language model architecture is as follows:

[0089] 1. Contains attention mechanism / multi-head attention mechanism;

[0090] 2. Contains Emebdding layer;

[0091] 3. Contains decode-only architecture;

[0092] 4. Contains efficiency improvement layer, such as layerNorm / droput, Gemma / V3 / Llama / Qwen3, etc.

[0093] Figure 2 It includes large model architecture that meets the scale rate, training on self-developed molecular science corpus, fine-tuning on self-developed corpus and design patterns, ideal screening and verification.

[0094] Among them, the training on the self-developed molecular science corpus involves the data set name QO2MOL, the tokenizer Byte-PairEncoding for SMILES, and the bit-by-bit pairing encoding based on the linear chemical language representation of SMILES.

[0095] Fine-tuning on self-developed corpus and design patterns includes stabilizer design patterns: Molecular Dipole Moment molecular dipole moment-driven stabilizer design, and prediction library matching entries: QO2MOL-dipole vector annotation.

[0096] Ideal screening and verification include SM-based Prediction Model prediction, Graph Network-based PageRank Recommendation recommendation, RDKit-based Synthetic Accessibility Analysis scoring, and DFT-based Simulation fine verification.

[0097] The generative language model is used for generation tasks. Combined with the tokenizer based on the Simplified molecular input line entry system (SMILES), the goal of “performing specific design tasks for stabilizer molecules” is achieved.

[0098] The self-developed QO2MOL dataset is used in combination with the dipole design scheme to realize the fine-tuning of the stable agent "vertical application scene" generative language model.

[0099] The fine-tuning process of the specific generative language model is shown as A1-A3.

[0100] A1: In the pre-training stage, the corpus for fine-tuning of the generative language model is obtained from the QO2MOL dataset; wherein the QO2MOL dataset at least includes SMILES expression and molecular property label; the molecular property label is a label associated with the dipole intensity scheme; the molecular property label at least includes molecular dipole intensity.

[0101] The corpus for fine-tuning of the generative language model is the QO2Mol dataset that performs well in the pre-training stage, and the QO2Mol dataset that performs well mainly refers to model loss, log loss, and test molecule synthesizability.

[0102] Molecular polarity and other properties can be used in experimental device research to design perovskite passivation molecular stabilizers, but there is no reliable method in the prior art for intelligent / data-driven / generative model-driven stabilizer design scheme. This is related to the lack of large-scale, available, and professional QO2MOL datasets, and also related to the maturity of the intelligent design route. Therefore, this scheme selects molecules with higher dipole intensity from QO2MOL as the corpus for fine-tuning of the generative language model that performs well.

[0103] In order to achieve the above-mentioned "perform specific design tasks of stabilizer molecules" goal, the survival language model is used to perform pre-training tasks on the self-developed QO2MOL dataset. The process of using the survival language model to perform pre-training tasks on the self-developed QO2MOL dataset is as follows:

[0104] The open source byte-based segmenter and the dictionary order commonly used for processing chemical linear language representation SMILES are used to establish a mapping relationship on the dataset;

[0105] The QO2MOL dataset already contains more than 20 million structurally reasonable and energy stable molecular data, and will be further expanded to more than 100 million molecular data, which provides good pre-training corpus for the training of the generative language model. These data are mapped into the input form of the generative language model by the segmenter, and the output data of the generative language model are decoded into SMILES.

[0106] The specific pre-training code is as follows:

[0107] python

[0108] deftrain_model():

[0109] filename = data_dir

[0110] checkpoint_path = base_checkpoint_path

[0111] checkpoint_path_prv = ". / checkpoints / YOUR_PATH"

[0112] model = GemmaForCausalLM.from_pretrained(f"{checkpoint_path_prv} / model", output_attentions=True)

[0113] # Trainer and Callbacks

[0114] checkpoint_cb = ModelCheckpoint(f"{checkpoint_path} / model / ")

[0115] early_stopping_ppl = EarlyStopping(

[0116] monitor="ppl_epoch", patience=4, min_delta=5e-3, check_finite=True,

[0117] stopping_threshold=1.1, divergence_threshold=hyperparams["vocab_size"] / 10,

[0118] verbose=True, mode="min", check_on_train_epoch_end=True )

[0120] trainer = Trainer(

[0121] strategy="ddp", accelerator="gpu", devices=1,

[0122] callbacks=[checkpoint_cb, early_stopping_ppl],

[0123] max_epochs = hyperparams["max_epochs"], min_epochs = hyperparams["min_epochs"],

[0124] ).

[0125] A2: processing the tokenizer into the input form of the generative language model; wherein the tokenizer contains a dictionary sequence and a mapping rule; the mapping rule is a mapping rule for converting the SMILES expression in the QO2MOL dataset into a molecular vector.

[0126] In converting the SMILES expression in the QO2MOL dataset into a molecular vector, i.e. the Toeknzier form, the conversion is performed by calling a language model architecture with generative inference capability.

[0127] A3: The corpus and input form of the fine-tuning of the generative language model are imported into the fine-tuning process of the generative language model.

[0128] In fine-tuning the generative language model, the QO2Mol dataset that performs well in the pre-training stage is first called, and a pre-set threshold is selected to load the selected high dipole strength dataset, i.e. the dataset of the fine-tuning task, which is processed into the input form of the language model by the tokenizer and imported into the fine-tuning process of the generative language model. In the fine-tuning process of the generative language model, a low linear algebra matrix basic property (Rank) trainer can be selected but is not necessary, and model gradient optimization is performed on the generative language model, including but not limited to an autoregressive algorithm-based model gradient optimization. The autoregressive algorithm-based model gradient optimization is a general solution for the generative language model.

[0129] Wherein, the pre-set threshold is set according to the actual situation, which is not limited in the present application.

[0130] The fine-tuned generative language model is a converged generative language model.

[0131] Wherein, the QO2Mol dataset that performs well mainly refers to model loss, log loss, and synthetic test molecule indicators.

[0132] Specific high dipole strength dataset, i.e. the dataset of the fine-tuning task, is as follows:

[0133] Plain Text;

[0134] NC(=O)NC=O;

[0135] CN1C=NN=C1;

[0136] NC1=COC(=O)N1;

[0137] O=C1C=CNC=C1;

[0138] O=C1C=CNC=N1;

[0139] O=C1NC=CC=N1;

[0140] N#CC1=CNC=C1;

[0141] N#CC1=CNC=N1;

[0142] O=CC1=NC=CN1;

[0143] O=CC1=CNC=N1.

[0144] S102: Construct a molecular vector corresponding to a preset number of stabilizer molecule candidates according to a pre-constructed prediction model, and obtain a selected candidate according to the molecular vector and a sorting algorithm (i.e., a sorting recommendation algorithm).

[0145] In S102, the fine-tuned generative language model is used to perform the stabilizer candidate generation task, and in the process of performing the stabilizer candidate generation task, a preset number of stabilizer molecule candidates are randomly generated.

[0146] The prediction model includes but is not limited to a supervised training machine learning model, a weakly supervised deep learning model, etc. The molecular features used to construct the prediction model include but are not limited to molecular fingerprints, molecular Coulomb matrices, molecular physical and chemical properties, molecular 3D structures, molecular potential energy data, etc.

[0147] The preset number of stabilizer molecule candidates are the recommended stabilizer candidates.

[0148] The converged generative language model is called, and the open source generative language model is pre-trained and fine-tuned with self-developed data sets, has a new form compared with the open source architecture, and can handle tasks that the open source architecture cannot handle. A preset number of stabilizer molecule candidates are generated.

[0149] The fine-tuned model is used to perform the stabilizer candidate generation task, and in the process of performing the stabilizer candidate generation task, a preset number of stabilizer molecule candidates are randomly generated by calling the fine-tuned generative language model. An example of using the fine-tuned generative language model to generate candidates is as follows:

[0150] sql

[0151] for _ in tqdm.tqdm(range(n_generated)):

[0152] Generate from " <s>" so that the next token is arbitrary.

[0153] smiles_start = torch.LongTensor([[tokenizer.bos_token_id]])

[0154] # Get generated token IDs.

[0155] generated_ids = model.generate(smiles_start,

[0156] max_length=hyperparams["max_length"],

[0157] do_sample=True, top_p=hyperparams["top_p"],

[0158] pad_token_id=tokenizer.eos_token_id)

[0159] # Decode the IDs into tokens and remove " <s>" and "< / s> ".

[0160] generated_smiles = tokenizer.decode(generated_ids[0],

[0161] skip_special_tokens=True)

[0162] generated_smiles_list.append(generated_smiles)。

[0163] The construction process of a specific prediction model is shown as B1-B9.

[0164] B1: Obtain an extended dataset of the QO2MOL dataset.

[0165] B2: Select a data loading method according to an open source machine learning tool.

[0166] The data loading method includes but is not limited to priority loading, offline loading, automatic loading, etc.

[0167] B3: Load the QO2MOL dataset and the extended dataset according to the data loading mode, to extract the effective molecular structure formula from the QO2MOL dataset, and extract the molecular dipole intensity label from the extended dataset.

[0168] B4: Divide the QO2MOL dataset and the extended dataset according to the preset proportion respectively, to obtain the training set, the test set and the validation set.

[0169] Wherein, the preset proportion is set according to the actual situation, and the present application does not make specific limitation.

[0170] B5: Obtain the molecular vector according to the molecular vector construction mode and the effective molecular structure formula.

[0171] Wherein, the molecular vector construction mode available at the present stage and in the future is selected.

[0172] B6: Input the molecular vector into the initial prediction model to obtain the output result of the initial prediction model.

[0173] In B6, the molecular vector is input into the learnable and trainable artificial intelligence model, i.e. the initial prediction model. In the example, the Sklearn tool is used as an example, but a wider range of implementation schemes can be used, such as various deep learning models based on neural networks, language models combined with word segmenters, etc. The output result of the trainable model is obtained by inputting the molecular vector.

[0174] Wherein, the initial prediction model is an initial prediction model that has not been trained, tested and verified.

[0175] B7: In any time period, obtain the loss function by the molecular dipole intensity label and the output result of the initial prediction model.

[0176] In B7, the label of QO2MOL is trained according to the loss function available in any time period, and the model parameters with low error and high generalization are selected for subsequent candidate screening.

[0177] B8: Train the initial prediction model by the loss function and the training set.

[0178] B9: Test the trained initial prediction model by the test set, and verify the tested initial prediction model by the validation set, to complete the process of constructing the prediction model.

[0179] The specific prediction model implementation example is as follows:

[0180] Plain Text

[0181] from sklearn.svm import SVR;

[0182] from sklearn.model_selection import train_test_split;

[0183] from sklearn.metrics import mean_squared_error;

[0184] svr = SVR(kernel='rbf',C=80, gamma=0.0105);

[0185] #SVR(C=80, gamma=0.0105).

[0186] Specifically, according to the pre-constructed prediction model, a preset number of molecular vectors corresponding to the stabilizer molecule candidates are constructed, and according to the molecular vectors and the sorting algorithm, a process of obtaining a selected candidate is obtained. The expected number of processes is shown as C1-C5.

[0187] C1: inputting the preset number of stabilizer molecule candidates into the pre-constructed prediction model to obtain the prediction model output result, and taking the prediction model output result as the property label of the candidate.

[0188] Wherein, the preset number is set according to the actual situation, and the present application is not limited.

[0189] C2: constructing a molecular vector corresponding to a preset number of stabilizer molecule candidates according to the property label of the candidate and the molecular annotation tool.

[0190] In C2, the output result of the prediction model with the preset number of stabilizer molecule candidates as input is obtained, that is, the preset number of stabilizer molecule candidates is input into the prediction model to obtain the output result of the prediction model, which is taken as the property label of the candidate. According to the property label of the candidate and other arbitrary time point publicly available molecular annotation tool, a description vector is constructed, and the description vector is taken as the input of the prediction model to become a molecular vector. The molecular vector is taken as the output result of the prediction model, which is called label.

[0191] Wherein, the label-scalar data is taken as a 0-dimensional vector.

[0192] C3: calculating the distance similarity between the molecular vectors corresponding to the preset number of stabilizer molecule candidates.

[0193] C4: constructing an adjacency matrix on the distance similarity through a recommendation algorithm, and constructing a graph network based on the adjacency matrix.

[0194] The visualization of the input adjacency matrix for the recommendation algorithm is as follows Figure 3 The similarity between each pair of molecules is intuitively presented by the side color wheel, and the higher the similarity, the closer to 1.

[0195] Figure 3 Among them, there is a side color wheel, intuitively presenting the similarity between each pair of molecules, and the higher the similarity, the closer to 1.

[0196] The 0-161 numerical label is the natural order number of different molecules, and has no other meaning except distinguishing molecules. On the basis of these data, the recommendation algorithm gives a recommendation score of different molecules.

[0197] C5: The ranking algorithm and molecular vector based on the graph network are used to obtain the expected number of selected candidates.

[0198] Among them, the ranking algorithm based on the graph network includes but is not limited to PageRank ranking algorithm.

[0199] The expected number can be 5, 10, etc. The expected number is set according to the actual situation, and the present application does not make specific limitations.

[0200] S103: Labeling the expected number of selected candidates by using a chemical informatics tool.

[0201] In S103, the recommended expected number of selected candidates is labeled by using a chemical informatics tool. The properties of the label include but are not limited to the synthesizability of the molecule. An example of calculating the synthesizability of the molecule is as follows:

[0202] Plain Text

[0203] from utils.cal_SA import sascorer

[0204] mol_ = Chem.MolFromSmiles(s_data)

[0205] sa_ = sascorer.calculateScore(mol_).

[0206] S104: Selecting target molecules satisfying a dipole strength threshold from the labeled expected number of selected candidates.

[0207] In S104, the expected number of target molecules satisfying the dipole strength threshold is selected from the labeled expected number of selected candidates by using a density functional theory framework, including but not limited to molecules satisfying the highest dipole strength threshold.

[0208] It should be noted that the target molecules are molecules with high tendency (such as molecules that are easier to synthesize).

[0209] Theoretical modeling and fine selection are performed from the expected number of labeled pre-selected candidates by the density functional theory framework, i.e. further sub-local molecular attention properties such as syntheticability, price cost, social economic factor scoring, etc. are scored to further select the target molecule with the highest dipole strength.

[0210] Among them, several molecules with the highest dipole strength are recommended as stabilizer configurations. The stabilizer configuration refers to the spatial structure and arrangement of the stabilizer molecule. The calculation results are shown in Table 1.

[0211] Table 1

[0212] SMILES Dipole Moment C(=O)NC(=O)NC(O) 8.704612 C(=O)NCC(F) 2.036424 C(=O)NNC(=O)C1CC1 6.317153

[0213] Among them, Dipole Moment is the dipole moment.

[0214] In summary, the present scheme gives an effective stabilizer, such as C(=O)NC(=O)NC(O). Its 3D configuration is shown in Figure 4 .The schematic diagram shows the simulation and verification of the physical and chemical process of the photostabilizer acting in the perovskite device. Figure 4

[0215] Figure 4 Among them, the red line in a indicates the molecule, and its SMILES expression is obtained by using the QO2MOL pre-training-fine-tuning of the generation formula language model. The green line in a indicates the iodine defect on the FAPbI3 halide perovskite surface. Figure 4 a is the SLAB model of the (100) surface of FAPbI3 used to represent the defect state of the interface of the photovoltaic device.

[0216] The first-principles calculation software package (Vienna Ab-initio Simulation Package, VASP) is used to obtain the free energy of the (100) surface of FAPbI3 when the iodine defect has been formed; and the free energy of the stabilizer (passivation) molecule in the same size cell; and the free energy after the stabilizer (passivation) molecule interacts with the surface iodine defect. The energy difference between the three states is shown in Figure 4 b, the combination process of the stabilizer molecule and the surface iodine defect is an exothermic process (the red line is the total energy of the two, and the total energy after passivation is lower, Figure 4 the green line in b is the absolute value of the heat released). The heat released (135.13) brings about the improvement of the stability of the surface iodine defect, which accounts for 27.75% compared to the energy of 486.95 before passivation. From the thermodynamic level, the effectiveness of the molecule passivation can be seen.

[0217] The physicochemical processes of the light stabilizer described in this invention in perovskite devices were simulated and verified using the VASP first-principles simulation tool. Density functional theory (DFT) data showed that the light stabilizer provided by this novel scheme can improve the stability of iodine defects on the perovskite surface by more than 27.75%.

[0218] exist Figure 4 c and Figure 4 In the d-squared diagram, electronic density of states and charge density distributions are used to analyze microscopic interactions. In the FAPbI3 semiconductor, electrons from Pb and I coordinate and occupy the conduction and valence bands, specifically manifested as the overlap of the centers of the Pb / I electron density distributions at the conduction and valence band edges. In the perovskite, electrons from oxygen atoms in the stabilizer occupy the conduction band edge in the density of states diagram, and their centers overlap with the Pb electron distribution at the conduction band edge. This means that the passivator molecules form an interaction similar to Pb-I near the conduction band. Figure 4 The charge distribution diagram of d in the figure uses the (010) section at the location of the iodine defect, which visually shows the void created in the charge distribution space after the formation of the iodine defect. The charge from the stabilizer fills this void. Furthermore, the polarity of the Pb charge distribution at the central position suggests that the central Pb atom interacts with the stabilizer molecule. In summary, based on a reasonable surface model and sufficient first-principles data, direct evidence is provided for the effectiveness of the molecule designed in this scheme.

[0219] S105: Using the target molecule as the target stabilizer configuration, and designing light stabilizers for perovskite photovoltaic devices based on the target stabilizer configuration.

[0220] The target stabilizer configuration is the recommended stabilizer configuration.

[0221] Stabilizer configuration refers to the spatial structure and arrangement of stabilizer molecules.

[0222] The purpose of this plan is:

[0223] 1. Due to the lack of a photovoltaic perovskite stabilizer process with a "data-AI model generation-AI screening-theoretical verification" model in existing technologies, one of the important factors is the lack of key domain datasets / scientific corpora. QO2MOL provides crucial support for addressing this issue.

[0224] 2. Language models under the scaling law effect have significant practical and industrial value in vertical applications within the scientific field. However, in the key issues of the aforementioned new energy photovoltaic perovskite industry, there is still a lack of feasible and effective solutions and cases driven by generative language models. Therefore, this solution provides a novel generative language model-driven design process to address the lack of feasible and effective solutions and cases driven by generative language models.

[0225] 3. This solution constructs an effective set of new solutions, including "self-developed dataset - self-developed material science large model molecular design module - standardized screening - industry-recognized theoretical verification", to help improve the stability iteration efficiency of the perovskite new energy device industry, providing a practical solution to reduce R&D cycle and cost.

[0226] The innovative aspects of this solution are:

[0227] 1. By combining novel technologies such as generative language models and hybrid expert MoE architecture, the significant level of intelligence brought about by the scaling law effect can be used to solve the industrial technology research and development problems of perovskite new energy devices.

[0228] 2. Using the self-developed QO2MOL molecular science corpus, a solution is provided for the above-mentioned problems (pre-training and fine-tuning of the generative model, and training of the predictive model).

[0229] 3. The light stabilizer candidates designed by combining the above process with molecular dipole strength (dipole rate, polarization intensity, etc., which describe the polar properties of molecules) were finally selected through sorting and recommendation algorithms, prediction model screening, and computational chemistry tools. The selected target molecules showed effective passivation and stability improvement effects on the surface iodine defects of perovskite devices.

[0230] This solution provides novel artificial intelligence tools to support the generation and selection of technical solutions for the iteration of light stabilizer technology in the perovskite new energy industry. It also provides a self-developed dataset, advanced generative language model pre-training and fine-tuning, and a hybrid expert architecture (MoE) for the research and development of light stabilizer technology in the perovskite new energy industry. Furthermore, it improves the stability of surface iodine defects that affect the energy conversion efficiency of organic-inorganic hybrid perovskites, with combined energy data showing an improvement of 27.75%. Electronic state density and charge distribution also verify that the stabilizer designed using this invention can chemically interact with surface iodine defects.

[0231] The beneficial effects of the embodiments of the present application are: adopting the self-developed QO2MOL molecular scientific corpus and the design scheme of the dual-driven perovskite photovoltaic device light stabilizer of the generated language model, through the design scheme of the dual-driven perovskite photovoltaic device light stabilizer, combining the molecular dipole strength such as dipole rate and polarization intensity to describe the polarity of the molecule, designing the stabilizer molecule candidate, and through the sorting algorithm, prediction model screening and calculation chemistry tool fine screening, finally selecting the molecule with the highest dipole strength, the selected molecule with the highest dipole strength shows effective passivation and stability improvement effect on the surface iodine defect of the perovskite device, and the molecule with the highest dipole strength is used as the target stabilizer configuration, and the perovskite photovoltaic device light stabilizer is designed through the target stabilizer configuration to improve the stability of the device microstructure.

[0232] Based on the above embodiments Figure 1 The disclosed design method of the perovskite photovoltaic device light stabilizer, the embodiments of the present application also correspondingly disclose a design system of the perovskite photovoltaic device light stabilizer, as shown in Figure 5 The design system of the perovskite photovoltaic device light stabilizer includes:

[0233] The first acquisition unit 501 is configured to execute a stabilizer candidate generation task by using the fine-tuned generated language model based on the QO2MOL data set and the dipole design scheme, and obtain a preset number of stabilizer molecule candidates.

[0234] The second acquisition unit 502 is configured to construct a molecular vector corresponding to the preset number of stabilizer molecule candidates according to a pre-constructed prediction model, and obtain an expected number of selected candidates according to the molecular vector and a sorting algorithm.

[0235] The marking unit 503 is configured to mark the expected number of selected candidates by using a chemical informatics tool.

[0236] The selection unit 504 is configured to select a target molecule satisfying a dipole strength threshold from the marked expected number of selected candidates.

[0237] The design unit 505 is configured to use the target molecule as a target stabilizer configuration, and design the perovskite photovoltaic device light stabilizer through the target stabilizer configuration.

[0238] Further, the first acquisition unit 501 of the fine-tuning process of the generated language model includes:

[0239] The first obtaining module is configured to obtain, in a pre-training stage, a corpus for fine-tuning of the generative language model from a QO2MOL dataset; wherein the QO2MOL dataset at least includes SMILES expressions and molecular property labels; the molecular property labels are labels associated with a dipole strength scheme; and the molecular property labels at least include molecular dipole strength;

[0240] The generating module is configured to process the tokenizer into an input form of the generative language model; wherein the tokenizer includes a dictionary order and a mapping rule; and the mapping rule is a mapping rule for converting the SMILES expressions in the QO2MOL dataset into molecular vectors;

[0241] The importing module is configured to import the corpus for fine-tuning of the generative language model and the input form into a fine-tuning process of the generative language model.

[0242] Further, the first obtaining unit 501 includes:

[0243] The executing module is configured to perform a stabilizer candidate generation task by the fine-tuned generative language model and the QO2MOL dataset and the dipole design scheme;

[0244] The random generating module is configured to randomly generate a preset number of stabilizer molecule candidates in the process of performing the stabilizer candidate generation task.

[0245] Further, the second obtaining unit 502 of the process of constructing the prediction model includes:

[0246] The second obtaining module is configured to obtain an extended dataset of the QO2MOL dataset;

[0247] The selecting module is configured to select a data loading mode according to an open-source machine learning tool;

[0248] The loading and extracting module is configured to load the QO2MOL dataset and the extended dataset according to the data loading mode, to extract valid molecular structural formulas from the QO2MOL dataset, and to extract molecular dipole strength labels from the extended dataset;

[0249] The dividing module is configured to divide the QO2MOL dataset and the extended dataset according to a preset proportion, respectively, to obtain a training set, a test set, and a validation set;

[0250] The third obtaining module is configured to obtain molecular vectors according to a molecular vector construction mode and the valid molecular structural formulas;

[0251] The fourth obtaining module is configured to input the molecular vectors into an initial prediction model to obtain an output result of the initial prediction model;

[0252] The fifth acquisition module is configured to acquire a loss function by the molecular dipole strength label and an output result of the initial prediction model within an arbitrary time period.

[0253] The training module is configured to train the initial prediction model by the loss function and the training set.

[0254] The test and verification module is configured to test the trained initial prediction model by the test set, and verify the tested initial prediction model by the verification set, so as to complete the process of constructing the prediction model.

[0255] Further, the second acquisition unit 502 comprises:

[0256] The sixth acquisition module is configured to input the preset number of stabilizer molecule candidates to the prediction model constructed in advance, obtain a prediction model output result, and take the prediction model output result as a property label of the candidate.

[0257] The construction module is configured to construct a molecular vector corresponding to the preset number of stabilizer molecule candidates according to the property label of the candidate and a molecular labeling tool.

[0258] The calculation module is configured to calculate a distance similarity between the molecular vectors corresponding to the preset number of stabilizer molecule candidates.

[0259] The construction module is configured to construct an adjacency matrix on the distance similarity by a recommendation algorithm, and construct a graph network based on the adjacency matrix.

[0260] The seventh acquisition module is configured to obtain an expected number of selected candidates by a sorting algorithm realized based on the graph network and the molecular vector.

[0261] Further, the selection unit 504 is specifically configured to select an expected number of target molecules from the labeled expected number of selected candidates by a density functional theory framework.

[0262] The beneficial effects of the embodiments of the application are as follows: the design scheme of the double-driven perovskite photovoltaic device light stabilizer generated by the self-developed QO2MOL molecular scientific corpus and the natural language model, by the design scheme of the double-driven perovskite photovoltaic device light stabilizer, in combination with the molecular dipole strength such as the dipole rate and the polarization strength to describe the polarity property of the molecule, the stabilizer molecule candidate is designed, the highest dipole strength molecule is finally selected by the sorting algorithm, the prediction model screening and the computational chemistry tool fine screening, the highest dipole strength molecule selected shows effective passivation and stability improvement effect on the surface iodine defect of the perovskite device, the highest dipole strength molecule is taken as the target stabilizer configuration, the perovskite photovoltaic device light stabilizer is designed by the target stabilizer configuration, so as to improve the stability of the device microstructure.

[0263] The embodiment of the present application further provides a storage medium comprising stored instructions, wherein the instructions, when executed, control a device where the storage medium is located to perform the design method of the perovskite photovoltaic device light stabilizer.

[0264] The embodiment of the present application further provides an electronic device, a structural schematic diagram of which is shown in Figure 6 The embodiment of the present application further provides an electronic device, a structural schematic diagram of which is shown in

[0265] For each method embodiment described above, in order to simply describe, each method embodiment is described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, some steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0266] It should be noted that each embodiment in the specification is described in a progressive manner, and each embodiment mainly explains the difference from other embodiments, and the same and similar parts between each embodiment can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0267] The steps in the method of each embodiment of the present application can be adjusted in sequence, combined and reduced according to actual needs.

[0268] Finally, it should be further noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or sequence between the entities or operations.

[0269] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0270] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.< / s>

Claims

1. A method of designing a photostabilizer for a perovskite photovoltaic device, characterized by, The method comprises: a generating language model fine-tuned through a QO2MOL dataset and a dipole design scheme is used to perform a stabilizer candidate generation task to obtain a preset number of stabilizer molecule candidates; a molecular vector corresponding to the preset number of stabilizer molecule candidates is constructed according to a pre-constructed prediction model, and an expected number of selected candidates is obtained according to the molecular vector and a sorting algorithm; the expected number of selected candidates is marked through a chemical informatics tool; a target molecule satisfying a dipole intensity threshold is selected from the expected number of selected candidates; the target molecule is used as a target stabilizer configuration, and a perovskite photovoltaic device light stabilizer is designed through the target stabilizer configuration.

2. The method of claim 1, wherein, The fine-tuning process of the generating language model comprises: in a pre-training stage, corpus for fine-tuning of the generating language model is obtained from the QO2MOL dataset; wherein the QO2MOL dataset at least comprises SMILES expressions and molecular property labels; the molecular property labels are labels associated with a dipole intensity scheme; the molecular property labels at least comprise molecular dipole intensity; a tokenizer is processed into an input form of the generating language model; wherein the tokenizer comprises a dictionary sequence and a mapping rule; the mapping rule is a mapping rule for converting the SMILES expressions in the QO2MOL dataset into a molecular vector; the corpus for fine-tuning of the generating language model and the input form are imported into the fine-tuning process of the generating language model.

3. The method of claim 1, wherein, The generating language model fine-tuned through the QO2MOL dataset and the dipole design scheme is used to perform the stabilizer candidate generation task to obtain the preset number of stabilizer molecule candidates, which comprises: the generating language model fine-tuned through the QO2MOL dataset and the dipole design scheme is used to perform the stabilizer candidate generation task; in the process of performing the stabilizer candidate generation task, a preset number of stabilizer molecule candidates are randomly generated.

4. The method of claim 1, wherein, The process of constructing the prediction model comprises: an extended dataset of the QO2MOL dataset is obtained; a data loading mode is selected according to an open-source machine learning tool; the QO2MOL dataset and the extended dataset are loaded according to the data loading mode to extract effective molecular structural formulas from the QO2MOL dataset and molecular dipole intensity labels from the extended dataset; the QO2MOL dataset and the extended dataset are respectively divided according to a preset ratio to obtain a training set, a test set and a validation set; a molecular vector is obtained according to a molecular vector construction mode and the effective molecular structural formulas; the molecular vector is input into an initial prediction model to obtain an output result of the initial prediction model; a loss function is obtained through the molecular dipole intensity labels and the output result of the initial prediction model within any time period; the initial prediction model is trained through the loss function and the training set; the trained initial prediction model is tested through the test set, and the tested initial prediction model is verified through the validation set to complete the process of constructing the prediction model.

5. The method of claim 1, wherein, The method comprises the following steps: inputting the preset number of stabilizer molecule candidates into a pre-constructed prediction model to obtain a prediction model output result, and taking the prediction model output result as a property label of the candidate; constructing a molecular vector corresponding to the preset number of stabilizer molecule candidates according to the property label of the candidate and a molecular labeling tool; calculating the distance similarity between the molecular vectors corresponding to the preset number of stabilizer molecule candidates; constructing an adjacency matrix on the distance similarity through a recommendation algorithm, and constructing a graph network based on the adjacency matrix; obtaining a desired number of selected candidates through a sorting algorithm realized based on the graph network and the molecular vector.

6. The method of claim 1, wherein, The method comprises the following steps: selecting a desired number of target molecules from the desired number of selected candidates marked.

7. A design system of a perovskite photovoltaic device light stabilizer characterized by, The system comprises: a first acquisition unit configured to execute a stabilizer candidate generation task through a generated language model fine-tuned by a QO2MOL dataset and a dipole design scheme to obtain a preset number of stabilizer molecule candidates; a second acquisition unit configured to construct a molecular vector corresponding to the preset number of stabilizer molecule candidates according to a pre-constructed prediction model, and obtain a desired number of selected candidates according to the molecular vector and a sorting algorithm; a marking unit configured to mark the desired number of selected candidates through a chemical informatics tool; a selection unit configured to select a target molecule satisfying a dipole intensity threshold from the desired number of selected candidates marked; a design unit configured to take the target molecule as a target stabilizer configuration, and design a perovskite photovoltaic device light stabilizer through the target stabilizer configuration.

8. The system of claim 7, wherein, The first acquisition unit of the fine-tuning process of the generated language model comprises: a first acquisition module configured to acquire a corpus for fine-tuning of the generated language model from the QO2MOL dataset in a pre-training stage; wherein the QO2MOL dataset at least comprises SMILES expressions and molecular property labels; the molecular property labels are labels associated with a dipole intensity scheme; and the molecular property labels at least comprise molecular dipole intensity; a generation module configured to process a tokenizer into an input form of the generated language model; wherein the tokenizer comprises a dictionary sequence and a mapping rule; and the mapping rule is a mapping rule for converting the SMILES expressions in the QO2MOL dataset into molecular vectors; an import module configured to import the corpus for fine-tuning of the generated language model and the input form into the fine-tuning process of the generated language model.

9. The system of claim 7, wherein, The first acquisition unit comprises: an execution module configured to execute a stabilizer candidate generation task through a generated language model fine-tuned by a QO2MOL dataset and a dipole design scheme; The random generation module is configured to randomly generate a preset number of stabilizer molecule candidates in a process of performing a stabilizer candidate generation task.

10. The system of claim 7, wherein, The second acquisition unit of the process of constructing the prediction model comprises: The second acquisition module is configured to acquire an extended dataset of the QO2MOL dataset; The selection module is configured to select a data loading manner according to an open source machine learning tool; The loading extraction module is configured to load the QO2MOL dataset and the extended dataset according to the data loading manner, to extract an effective molecular structural formula from the QO2MOL dataset, and to extract a molecular dipole intensity label from the extended dataset; The division module is configured to divide the QO2MOL dataset and the extended dataset according to a preset proportion, respectively, to obtain a training set, a test set and a verification set; The third acquisition module is configured to acquire a molecular vector according to a molecular vector construction manner and the effective molecular structural formula; The fourth acquisition module is configured to input the molecular vector into an initial prediction model to obtain an output result of the initial prediction model; The fifth acquisition module is configured to acquire a loss function by using the molecular dipole intensity label and the output result of the initial prediction model within any time period; The training module is configured to train the initial prediction model by using the loss function and the training set; The test verification module is configured to test the trained initial prediction model by using the test set, and to verify the tested initial prediction model by using the verification set, so as to complete the process of constructing the prediction model.

Images