Method for predicting advantages and disadvantages of acetylene hydrochlorination catalyst ligand based on deep neural network

Abstract

The invention discloses an acetylene hydrochlorination catalyst ligand quality prediction method based on a deep neural network, and belongs to the field of material informatics. The method comprises the following steps of (1) collecting existing public literatures to establish a database, (2) extracting an SMILES character string, a reaction condition and a performance improvement rate of a ligand and constructing a tensor matrix, (3) adopting Rul as a hidden function to carry out back propagation training and save a model result, and (4) inputting an unknown ligand and the reaction condition to obtain an effect ranking. The invention further provides a computer readable storage medium storing the method and a device composed of the medium. According to the method, reaction conditions are increased, and the accuracy of predicting the advantages and disadvantages of the catalyst ligand in the acetylene hydrochlorination reaction is improved.

Description

Technical Field

[0001] The present invention relates to a method for predicting the quality of acetylene hydrochlorination reaction catalyst ligands based on a deep neural network, in particular to a method that combines the knowledge of catalysis and deep neural networks and implements an artificial intelligence-assisted screening of the quality of acetylene hydrochlorination reaction catalyst ligands, belonging to the technical field of materials informatics. Background Art

[0002] Polyvinyl chloride (PVC) is a very important chemical raw material, which is mainly used in the production process of plastics. Polyvinyl chloride is a raw material with great demand, and polyvinyl chloride is mainly polymerized from vinyl chloride as raw material. Therefore, the synthesis of vinyl chloride is the most important process in the production process of polyvinyl chloride, and vinyl chloride is mainly made from the hydrochlorination reaction of acetylene. In the chemical production process, catalysts are indispensable, and efficient acetylene hydrochlorination catalysts are particularly important. In the chemical production process, the traditional practice is to use a catalyst containing HgCl2 to hydrochlorinate acetylene. Hg, as a heavy metal element, can cause great harm to the human body and the natural environment. On top of this environmental problem, finding a catalyst that is equally efficient and clean and pollution-free has become the most important issue at present. At present, the research on mercury-free catalysts is mainly based on the following four directions: catalysts based on ruthenium, platinum, gold and copper. This patent mainly predicts high-performance acetylene hydrochlorination catalysts based on the above four substrates through a machine learning model based on a deep neural network algorithm.

[0003] The performance of catalysts is usually affected by multiple factors, such as the size of atoms, the valence state of elements, and the structure of catalysts. In the design and prediction process of catalysts, the traditional method is usually used, that is, to screen out efficient mercury-free hydrochlorination catalysts through experiments. In this process, a lot of time and financial costs are usually consumed. With the rapid development of current science and technology, artificial intelligence technology has been greatly promoted. Artificial intelligence has been applied to various fields and is an excellent tool for assisting scientific research. This patent thinks of training a machine learning model to assist our efficient mercury-free hydrochlorination catalyst design and prediction process. Summary of the invention

[0004] The technical problem solved by the present invention is: a prediction method for the quality of ligands of acetylene hydrochlorination reaction catalysts based on deep neural networks is proposed. A model is established based on existing papers and experimental data based on deep neural networks (DNNs) to obtain a new prediction model, which is named the LDNN model, so as to predict the conversion rate of catalysts with unknown ligands added, and guide catalyst synthesis with the help of the output results of the model. The method is as follows: first, various relevant literatures are read, and the chemical ligands obtained therefrom are constructed into a ligand database. Then, a novel prediction model for the conversion rate of acetylene hydrochlorination catalysts with unknown ligands added is proposed, which combines feature extractors and DNN algorithms. The model can calculate the ECFP fingerprint of the relevant ligands according to the composite SMILES string, and then combine the tensor matrix splicing with specific reaction conditions to make a directional prediction of the conversion rate, thereby improving the accuracy of the prediction.

[0005] In order to solve the above technical problems, the technical solution proposed in the present invention is: a method for predicting the quality of ligands of acetylene hydrochlorination catalyst based on deep neural network, the method comprising the following steps:

[0006] (1) Collect and extract data from literature as data sets for model training; catalyst ligands, reaction conditions and ligand properties are used to train the model;

[0007] (2) Extract the composite SMILES string of the ligand molecule and the reaction conditions and concatenate the tensor matrix; each atom and chemical bond in the composite SMILES string is a word, obtain SMILES through public information, use ECFP (Extended Connectivity Fingerprint) technology to collect features, combine reaction conditions, train iteratively, and use Reul function;

[0008] (3) Input the tensor matrix into the model developed based on the deep neural network for training;

[0009] (4) Input the composite SMILES string of the unknown performance ligand and the reaction conditions to obtain the results of the ligand quality;

[0010] The reaction conditions and the ligand composite SMILES string are concatenated and a tensor matrix is ​​constructed, which is input into the model as an input variable. The trained LDNN model is used to embed the ligand composite SMILES string into the tensor and output the quality of the catalyst.

[0011] Preferably, the input modules are divided into three categories, a buffer layer and a hidden layer are added to each module, and finally the BCEloss loss function is used for gradient improvement.

[0012] Preferably, information extracted from the molecular library is input into the model, combined with the reaction conditions, to predict the performance of the unknown ligand.

[0013] Preferably, the method comprises the following steps:

[0014] Step 1: Establish a data set and collect information on known catalyst ligands from the literature. The molecular structure of the ligand and the corresponding composite SMILES string, temperature, acetylene space velocity, central metal content, HCl ratio, and conversion rate are recorded in the database;

[0015] Step 2: Define the catalyst conversion rate improvement

[0016] Conversion rate improvement rate (C l ), specifically , C e is the conversion rate of the catalyst control sample without ligand under the same conditions, C r is the conversion rate of the catalyst with added ligand;

[0017] Step 3: Convert the molecular structure of the ligand into an ECFP fingerprint

[0018] The composite SMILES string of each ligand is converted into a 256-bit binary vector called "ECFP fingerprint". For the SMILES string of known ligands, a search method is used to map each functional group to a molecular fingerprint composed of 0 and 1 from the RDKit open source library, and then combine them into a 256-bit molecular fingerprint. Different sequences of 0 and 1 represent different functional groups, thus reflecting the influence of different ligand structures.

[0019] Step 4: Construct the feature tensor matrix

[0020] Each ligand collected has its own 256-bit molecular fingerprint. The fingerprint data of all ligands are integrated into a tensor matrix. Each row of the matrix represents a ligand and each column is a feature bit.

[0021] The experimental conditions temperature and acetylene space velocity are converted into vectors and added to the characteristic matrix;

[0022] Again, the HCl ratio and central metal content experimental conditions were converted into vectors and added to the feature matrix;

[0023] Step 5: Build and tune the model architecture

[0024] Using the DNN algorithm, since the tensors containing airspeed and temperature have too much difference in magnitude, they are partitioned separately as separate input modules, and a buffer layer is added as the first layer, which is conducive to reducing the difference between different variables; for the three different input modules, the corresponding function Reul is used to convert them into three vectors, and they are spliced ​​into a matrix and used to correspond to a 1*1 tensor using the Reul function again. The BCEloss function is used to train the parameters of the corresponding function with the gradient boosting method, and the conversion rate is predicted;

[0025] Step 6: Model training and evaluation

[0026] The simplified feature matrix group was input into LDNN and training continued. The target variable was the conversion rate. Through model training, the training step size was adjusted to 0.001 and the number of training times was adjusted to 2000. The model gradually learned the relationship between ligand structure, experimental conditions and conversion rate.

[0027] At the same time, in order to prevent overfitting, the dropout function was used to discard 12% of the data each time, and the optimal value of each training parameter was determined after trial and error. After the training was completed, the model was evaluated using the validation set. In the end, the model's prediction accuracy for different ligands was less than 5%, with a probability of more than 80%.

[0028] Step 7: Predict the conversion rate of unknown ligands

[0029] After training, the model can be used to predict the conversion of a new ligand, ligand X; RDKit is used to generate a 256-bit fingerprint of ligand X, which is then merged into a feature matrix with the corresponding experimental conditions; after inputting its feature vector into the model, the model outputs the predicted conversion improvement rate; this output can help determine whether ligand X has a good catalytic effect under the reaction conditions.

[0030] The method for predicting the quality of ligands of acetylene hydrochlorination catalyst based on deep neural network is characterized in that the ligands are screened out.

[0031] As a catalyst for acetylene hydrochlorination.

[0032] A computer-readable storage medium having a computer program capable of running the method for predicting the quality of acetylene hydrochlorination reaction catalyst ligand based on a deep neural network as claimed in claim 1.

[0033] A device for predicting the quality of ligands of acetylene hydrochlorination reaction catalysts based on a deep neural network, the device being equipped with the computer-readable storage medium of claim 6.

[0034] The method comprises the following steps:

[0035] (1) Construct a database of relevant chemical ligands for model training; (2) Extract the ligands and convert them into 256-bit ECFP fingerprints and separate them bit by bit, and then construct a tensor matrix based on the reaction conditions; (3) Input the tensor matrix as a preliminary feature vector into the DNN, where each column in the tensor matrix is ​​a feature, and set the importance attribute of each feature to the model performance. Use the database data to preliminarily train the DNN model to obtain the importance parameter of each feature, sort them in descending order, and delete the features with relatively low importance in proportion to reduce the dimension and prevent overfitting; (4) Construct the environmental variables airspeed, temperature and HCL ratio into a tensor matrix of environmental variables, and merge it with the tensor matrix after dimensionality reduction to obtain a complete feature matrix; (5) Input the feature matrix into the DNN to train the model and evaluate its performance.

[0036] First, we collected a large amount of chemical ligand data by consulting various relevant literature. Each ligand needs to include its molecular structure, properties, and conversion rate in the acetylene hydrochlorination reaction. The ligand information was obtained through literature surveys, laboratory data collection, and the public chemical database PubChem.

[0037] Furthermore, the chemical ligands are feature extracted by ECFP (Extended Connectivity Fingerprint) technology. ECFP can effectively represent the topological information of the molecule and convert the chemical structure of the ligand into a fixed-length vector representation. The 256-bit ECFP fingerprint of each ligand is extracted using the chemical informatics tool RDKit. The neighborhood information of atoms or groups in the molecule is extracted through an iterative algorithm and constructed into a series of binary bit representations to generate a fixed-length molecular fingerprint. The extracted 64-bit ECFP fingerprint is converted into a tensor matrix. The ECFP fingerprint of each ligand is a 64-dimensional binary vector. When constructing the tensor matrix, the fingerprint of each ligand will become a row vector of a matrix. When there are multiple ligands, these row vectors will form a two-dimensional tensor, in which each row represents the fingerprint of a ligand.

[0038] Furthermore, the tensor matrix obtained in the previous step is used as the feature vector input, and the conversion rate is used as the target output to train an LDNN model. As the training progresses, the LDNN model uses the gradient boosting method to back-propagate and optimize the corresponding function of the hidden layer in each iteration, adjusting the amplitude and number of training each time. In addition, the influence of reaction conditions is also considered.

[0039] Furthermore, the experimental environmental variables closely related to the acetylene hydrochlorination reaction, such as space velocity, temperature and HCL ratio, are selected. These variables will affect the catalyst reaction performance accordingly, so it is necessary to consider these factors in the model. The environmental variables are converted into a tensor matrix, and the value of each of them is used as a column in the matrix to construct a two-dimensional tensor of environmental variables. The tensor matrix of environmental variables is merged with the tensor matrix after dimensionality reduction by column to obtain a complete feature matrix.

[0040] Furthermore, the complete feature matrix obtained in the previous step is used as input, and the conversion rate is used as the target variable to be input into the DNN model for training. During the training process, the BCEloss function is used to train the model using the gradient boosting method.

[0041] Beneficial effects:

[0042] The present invention uses a deep-learning neural network algorithm to train a model. The main advantage of deep-learning neural networks is that they can fit and predict when the flux is small, and have better performance when the order of magnitude difference is large. This type of model can greatly reduce the workload and has a high accuracy rate. We have used this algorithm to design and predict multiple high-performance mercury-free hydrochlorination catalysts, and have shown high accuracy and catalytic performance.

[0043] Existing catalyst conversion rate prediction methods (DFT models, etc.) are usually based on the solution of the Schrödinger equation, which is too complicated and requires a lot of computing resources. The universal screening method of small models using artificial intelligence blurs and blackboxes the intermediate process, highlights the input and output, and for the completed training model, the prediction of a single ligand only takes tens of milliseconds, greatly saving computing costs.

[0044] Existing catalyst conversion rate prediction methods usually focus on data features from a single source, such as modeling based only on the molecular structure information of the ligand or a single experimental environment variable. The defect of such a model is that it ignores the interaction between chemical ligands and experimental environment variables and their combined influence on the catalytic reaction performance. This model combines the ECFP fingerprint of the ligand with the experimental environment variables to predict the catalytic performance. In this way, the model not only takes into account the molecular structure of the ligand during the catalyst preparation process, but also can make dynamic predictions based on changes in experimental conditions, thereby more comprehensively capturing the factors affecting catalyst performance. This fusion of multiple features greatly improves the prediction accuracy of the model, especially in the face of unknown ligands and arbitrarily changing experimental conditions.

[0045] Many traditional methods use simple molecular descriptors or features to represent ligands. These methods usually fail to effectively capture the complexity of ligand structures, resulting in poor model performance. The present invention uses ECFP fingerprints to extract features of organic ligands and converts them into 256-bit binary vectors. By converting the ECFP fingerprint of each ligand into a tensor matrix and inputting it into the model, the present invention can efficiently express the complex information of the molecular structure. This tensor matrix form can capture higher-order and more hidden relationships in molecules, thereby improving the model's ability to represent ligands and prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described below in conjunction with the accompanying drawings.

[0047] Figure 1 It is a schematic diagram of the present invention DETAILED DESCRIPTION

[0048] Example 1

[0049] The technical solution of the present invention is further explained below through embodiments in combination with the accompanying drawings.

[0050] Step 1: Create a dataset

[0051] First, collect information about known catalyst ligands from the literature. For example, a catalyst ligand A was found, which carried out acetylene hydrochlorination reaction at 210°C and HCl ratio of 1.10 with a conversion rate of 70%. This information (ligand molecular structure and corresponding composite SMILES string, temperature, acetylene space velocity, central metal content, HCl ratio, conversion rate) was recorded in the database.

[0052] Step 2: Define the catalyst conversion rate improvement

[0053] In order to balance the impact of different batches of activated carbon from different manufacturers on the conversion rate, the conversion rate improvement rate (C l ) is used to describe, specifically C l = (C e is the conversion rate of the catalyst control sample without ligand under the same conditions, C r is the conversion rate of the catalyst with added ligand)

[0054] Step 3: Convert the molecular structure of the ligand into an ECFP fingerprint

[0055] In order to let the model "recognize" the structure of the ligand, the composite SMILES string of each ligand is converted into a 256-bit binary vector called "ECFP fingerprint". For the SMILES string of the known ligand, a search method is used to correspond each functional group to a molecular fingerprint composed of 0 and 1 from the RDKit open source library, and then combine them into a 256-bit molecular fingerprint. Different sequences of 0 and 1 represent different functional groups, thus reflecting the influence of different ligand structures.

[0056] For example, the composite SMILES string of ligand A is C=CBr, which is retrieved in the RDKit library and converted into the following 256-bit binary fingerprint: [1, 1, 0, 1, 0, 0, 1, ...]. Each bit in this string of data represents a detail of the molecular structure, such as whether it contains certain specific chemical bonds or groups. For subsequent model training, this fingerprint becomes the "identity card" of the ligand.

[0057] Step 4: Construct the feature tensor matrix

[0058] Assume that each ligand collected has its own 256-bit molecular fingerprint. Integrate the fingerprint data of all ligands into a tensor matrix. Each row of the matrix represents a ligand, and each column is a feature bit. For example, the first two rows can be like this:

[0059] | Ligand number | Position 1 | Position 2 | Position 3 | Position 4 | Position 5 | ... |

[0060] |----------|-----|-----|-----|-----|-----|-----|

[0061] | A | 1 | 1 | 0 | 0 | 1 | ... |

[0062] | B | 0 | 1 | 0 | 0 | 1 | ... |

[0063] Under this matrix structure, the molecular structural characteristics of all ligands are clearly visible, making it convenient for subsequent input into the model.

[0064] Then the experimental conditions are added. The conversion rate of acetylene hydrochlorination is not only affected by the ligand structure, but also closely related to the experimental conditions (such as temperature and space velocity). In order to improve the prediction accuracy, the experimental conditions (temperature and space velocity) are converted into vectors and added to the feature matrix.

[0065] Assume that the experimental temperature of ligand A is 210℃ and the space velocity is 340h -1 , these values ​​can be encoded as additional feature columns, such as:

[0066] | Ligand number| Temperature| Space velocity|

[0067] |----------|------|----------|

[0068] | A | 210 | 340 |

[0069] | B | 250 | 680 |

[0070] In addition, due to the large difference in magnitude, in order to prevent other experimental conditions (HCl ratio and central metal content) that affect the conversion rate of acetylene hydrochlorination reaction from being masked by the large magnitude changes of features, these features were separated separately and formed into a new matrix for separate input.

[0071] Again, the experimental conditions such as HCl ratio and central metal content are converted into vectors and added to the feature matrix. Assuming that the HCl ratio of ligand A is 1.10 and the central metal content is 1%, these values ​​are encoded as additional feature columns, such as:

[0072] | Ligand number | HCl ratio | Central metal content |

[0073] |----------|---------|------------------|

[0074] | A | 1.10 | 1% |

[0075] | B | 1.15 | 0.5% |

[0076] In this way, the feature matrix group contains not only the molecular structure information of the ligand, but also different experimental conditions, based on which the model can take into account the differences in reaction performance caused by environmental changes.

[0077] Step 5: Build and tune the model architecture

[0078] After deciding to use the DNN algorithm, considering that the tensors containing airspeed and temperature have too many orders of magnitude differences, they are partitioned separately as separate input modules, and a buffer layer is added as the first layer, which is conducive to reducing the differences between different variables. For the three different input modules, the corresponding function Reul is used to convert them into three vectors, and they are spliced ​​into a matrix and used to correspond to the Reul function again to become a 1*1 tensor. The BCEloss function is used to train the parameters of the corresponding function with the gradient boosting method, and the conversion rate is predicted. In addition, for these four corresponding functions, each function has two different options (linear and nonlinear), so it is important to ensure the adaptability of these functions.

[0079] Step 6: Model training and evaluation

[0080] The simplified feature matrix group was input into LDNN and training continued. The target variable was the conversion rate. Through model training, the training step size (0.01, 0.001, 0.0001) and the number of training times (500, 1000, 2000, 10000) were adjusted, and the model gradually learned the relationship between ligand structure, experimental conditions and conversion rate.

[0081] At the same time, in order to prevent overfitting, the dropout function is used to discard some data each time (6%, 12%, 24%, 48%), and the best value of each training parameter is determined after trial and error. After training, the model is evaluated using the validation set.

[0082] Assume that the conversion rate of the catalyst prepared by ligand A is predicted to be 70%, but the actual conversion rate is 50%, then the error is 10%. Through this evaluation, the model parameters are adjusted and determined as follows: (the corresponding function is Reul, the training step is 0.001, the number of training times is 2000, the dropout function parameter is set to 0.12, and the structure is as follows Figure 1 As shown in Figure 2), the probability that the final model’s prediction accuracy for different ligands was less than 5% reached more than 80%.

[0083] Step 7: Predict the conversion rate of unknown ligands

[0084] After training, the model can be used to predict the conversion of new ligands. Suppose there is a new ligand X, whose composite SMILES string is C[P+]. Use RDKit to generate a 256-bit fingerprint of ligand X, and then merge it with the corresponding experimental conditions into the feature matrix. For example, assuming the experimental conditions of X are temperature 210℃ and HCl ratio 1.10, after inputting its feature vector into the model, the model output predicts a conversion rate improvement of 70%. This output result can help determine whether ligand X has a good catalytic effect under the reaction conditions and whether it is worth further verification in the experiment.

[0085] More than 8000 ligands were collected from https: / / www.chemsrc.com / and converted into SMLILES strings, with the central metal content set to 1%, the HCl ratio set to 1.15, and the acetylene space velocity set to 680 h -1 , the reaction temperature was set to 180 degrees Celsius, and the model was brought in for prediction. About 100 ligands were obtained with excellent prediction results. After removing factors such as toxicity, danger, and price, 2 ligands were selected from the remaining ligands for verification.

[0086] Step 8: Prepare the catalyst

[0087] The ligand predicted to have good effect under the given reaction conditions in step 7 is used for impregnation method to prepare a catalyst with a central metal content of 1% Ru. The catalyst performance is tested experimentally using the same reaction conditions. At the same time, a catalyst with a central metal content of 1% Ru without adding ligand is prepared under the same conditions as a control to determine the improvement rate of the catalyst conversion rate after modification.

[0088] Step 9: Experimental Testing

[0089] For the predicted result C l = 0.71 ligand (CAS No.: 141556-45-8) and blank control were tested and the conclusion was drawn: r =88%, C e = 61%, calculate C l =0.69, the effect is good.

[0090] For the predicted result C l = 0.80 ligand (CAS No.: 79917-90-1) and blank control were tested and the conclusion was drawn: r =94%, C e = 61%, calculate C l =0.84, the effect is good.

[0091] The present invention is not limited to the specific technical solutions described in the above embodiments, and all technical solutions formed by equivalent replacement are within the protection scope required by the present invention.

Claims

1. A method for predicting the quality of catalyst ligands for acetylene hydrochlorination reaction based on deep neural network, characterized in that: The method comprises the following steps: (1) collecting and extracting data from literature as a data set for training a model; catalyst ligands, reaction conditions and ligand properties are used to train the model; the reaction conditions are acetylene space velocity, temperature, central metal content and HCl ratio; (2) Extract the composite SMILES string of the ligand molecule and the reaction conditions and concatenate the tensor matrix; each atom and chemical bond in the composite SMILES string is a word, obtain SMILES through public information, use ECFP (Extended Connectivity Fingerprint) technology to collect features, combine reaction conditions, train iteratively, and use Reul function; (3) Input the tensor matrix into the model developed based on the deep neural network for training; (4) Input the composite SMILES string of the unknown performance ligand and the reaction conditions to obtain the results of the ligand quality; The reaction conditions and the ligand composite SMILES string are concatenated and a tensor matrix is ​​constructed, which is input into the model as an input variable. The trained LDNN model is used to embed the ligand composite SMILES string into the tensor and output the quality of the catalyst.

2. The method for predicting the quality of acetylene hydrochlorination catalyst ligands based on deep neural network according to claim 1, characterized in that: It is divided into three types of input modules, namely, ligand, acetylene space velocity and temperature, central metal content and HCl ratio. A buffer layer and a hidden layer are added to each module, and finally the BCEloss loss function is used for gradient improvement.

3. The method for predicting the quality of acetylene hydrochlorination catalyst ligands based on deep neural network according to claim 1, characterized in that: The information pulled from the molecular library is fed into the model and combined with the reaction conditions to predict the performance of the unknown ligand.

4. The method for predicting the quality of acetylene hydrochlorination catalyst ligands based on deep neural network according to claim 1, characterized in that: The method comprises the following steps: Step 1: Establish a data set and collect information on known catalyst ligands from the literature. The molecular structure of the ligand and the corresponding composite SMILES string, temperature, acetylene space velocity, central metal content, HCl ratio, and conversion rate are recorded in the database; Step 2: Define the catalyst conversion rate improvement Conversion rate improvement rate (C l ), specifically C l = , C e is the conversion rate of the catalyst control sample without ligand under the same conditions, C r is the conversion rate of the catalyst with added ligand; Step 3: Convert the molecular structure of the ligand into an ECFP fingerprint The composite SMILES string of each ligand is converted into a 256-bit binary vector called "ECFP fingerprint". For the SMILES string of known ligands, a search method is used to map each functional group to a molecular fingerprint composed of 0 and 1 from the RDKit open source library, and then combine them into a 256-bit molecular fingerprint. Different sequences of 0 and 1 represent different functional groups, thus reflecting the influence of different ligand structures. Step 4: Construct the feature tensor matrix Each ligand collected has its own 256-bit molecular fingerprint. The fingerprint data of all ligands are integrated into a tensor matrix. Each row of the matrix represents a ligand and each column is a feature bit. The experimental condition temperature and acetylene space velocity are converted into vectors and added to the characteristic matrix; Again, the HCl ratio and central metal content experimental conditions were converted into vectors and added to the feature matrix; Step 5: Build and tune the model architecture Using the DNN algorithm, since the tensors containing airspeed and temperature have too much difference in magnitude, they are partitioned separately as separate input modules, and a buffer layer is added as the first layer, which is conducive to reducing the difference between different variables; for the three different input modules, the corresponding function Reul is used to convert them into three vectors, and they are spliced ​​into a matrix and used to correspond to a 1*1 tensor using the Reul function again. The BCEloss function is used to train the parameters of the corresponding function with the gradient boosting method, and the conversion rate is predicted; Step 6: Model training and evaluation The simplified feature matrix group is input into LDNN and training continues. The target variable is the conversion rate. Through model training, the corresponding function is Reul, the training step size is adjusted to 0.001, and the number of training times is 2000. The model gradually learns the relationship between ligand structure, experimental conditions and conversion rate. At the same time, in order to prevent overfitting, the dropout function is used to discard some data 12% each time, and the best value of each training parameter is determined after trial and error; After training, the model was evaluated using the validation set. The probability that the difference between the predicted conversion rate and the actual conversion rate of different ligands was less than 5% reached more than 80%; Step 7: Predict the conversion rate of unknown ligands After training, the model can be used to predict the conversion of a new ligand, ligand X; RDKit is used to generate a 256-bit fingerprint of ligand X, which is then merged into a feature matrix with the corresponding experimental conditions; after inputting its feature vector into the model, the model outputs the predicted conversion improvement rate; this output can help determine whether ligand X has a good catalytic effect under the reaction conditions.

5. The method for predicting the quality of acetylene hydrochlorination catalyst ligands based on deep neural network according to claim 1, characterized in that: Screening out ligands As a modifier for acetylene hydrochlorination catalyst.

6. A computer-readable storage medium, characterized in that: The storage medium has a computer program, and the computer program is capable of running the method for predicting the quality of acetylene hydrochlorination reaction catalyst ligands based on deep neural network as described in claim 1.

7. A prediction device for the quality of acetylene hydrochlorination catalyst ligands based on deep neural networks, characterized in that: The device is equipped with the computer-readable storage medium according to claim 6.

Images