A neural network prediction-based lactic acid polymerization process

Abstract

The application discloses a lactic acid polymerization process based on neural network prediction, and relates to prediction of lactic acid polymerization conditions. The process adopts a superimposed shadow sampling uniform design method, controls lactic acid polymerization reaction time, catalyst content and reaction temperature, obtains lactic acid polymers with different molecular weights, and trains and predicts a neural network by using obtained test data. The predicted data is in good agreement with the test data. The application greatly simplifies the test exploration process, saves test time and cost, and has wide application prospects in the field of biological chemical synthesis.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biological chemical industry, and particularly relates to a lactic acid polymerization process based on neural network prediction. BACKGROUND

[0002] Polylactic acid is a polyester high-molecular bio-based polymer material synthesized by polymerization of lactic acid, and has important applications in various application fields. In recent years, polylactic acid is mainly synthesized by lactide ring-opening polymerization. Although the lactide ring-opening polymerization method can obtain polylactic acid with a large molecular weight, the harsh reaction conditions and low lactide yield result in high cost of the produced polylactic acid, which limits the wide application of polylactic acid. The molecular weight of polylactic acid obtained by lactic acid polycondensation is lower than that obtained by lactide ring-opening polymerization, but the raw material lactic acid is widely available and low in price, so it is expected to become an important means to obtain polylactic acid by lactic acid polycondensation. However, the reaction conditions in the polycondensation process are complex, and a large number of experiments need to be carried out.

[0003] The use of superimposed sampling uniform design method and neural network can optimize and predict the reaction conditions of lactic acid direct polycondensation method, simplify the experimental exploration process, but there is no related report. SUMMARY

[0004] In view of the above problems existing in the prior art, the application aims to provide a lactic acid polymerization process based on neural network prediction, which reduces the number of experiments, saves experiment time and cost, and obtains a better lactic acid polymerization process.

[0005] The lactic acid polymerization process based on neural network prediction has the characteristics that the catalyst is put into the lactic acid liquid for heating reaction, reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as control independent variables, and the molecular weight of polylactic acid is used as a dependent variable. The combination scheme of the three factor control variables is used to carry out a reaction experiment for synthesizing polylactic acid, and experimental data of a vector composed of the independent variables and the dependent variable are obtained. The neural network has a three-layer neural network structure, the model algorithm used is Levenberg-Marquardt (hereinafter referred to as L-M algorithm), reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as input variables, and the molecular weight of polylactic acid is used as an output variable. A model based on neural network prediction is constructed, the model is used to train the above obtained experimental data, the corresponding relationship between the input variables and the output variables is obtained, and the trained model can predict the molecular weight of polylactic acid synthesized by the combination scheme of the three input variables.

[0006] The lactic acid polymerization process based on neural network prediction has the characteristics that the catalyst is put into the lactic acid liquid for heating reaction, reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as control independent variables, and the molecular weight of polylactic acid is used as a dependent variable. The combination scheme of the three factor control variables is used to carry out a reaction experiment for synthesizing polylactic acid, and experimental data of a vector composed of the independent variables and the dependent variable are obtained. The neural network has a three-layer neural network structure, the model algorithm used is Levenberg-Marquardt (hereinafter referred to as L-M algorithm), reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as input variables, and the molecular weight of polylactic acid is used as an output variable. A model based on neural network prediction is constructed, the model is used to train the above obtained experimental data, the corresponding relationship between the input variables and the output variables is obtained, and the trained model can predict the molecular weight of polylactic acid synthesized by the combination scheme of the three input variables.

[0007] S1: using the superposition sampling uniform design method, carrying out the experimental design of three factors, determining the synthesis conditions of the direct polycondensation method of polylactic acid by controlling the reaction temperature, reaction time and catalyst addition amount three factors, measuring the molecular weight of the synthesized polylactic acid, obtaining the experimental data about the vector composed of the three factors and the molecular weight of polylactic acid;

[0008] S2: constructing a three-layer neural network structure, the neural network structure for predicting the molecular weight of polylactic acid is as shown in Figure 1 The input layer contains 3 nodes, corresponding to 3 influencing factors, namely reaction temperature x1, reaction time x2 and catalyst addition amount x3; the number of nodes in the hidden layer is determined according to the Kolmogorov theorem, that is, the number of nodes in the input layer is n, then the number of nodes in the hidden layer is 2n+1, that is, 7; the output layer contains one node, that is, the molecular weight Y of polylactic acid. According to the input variables x1, x2 and x3, the connection weight value between the input layer and the hidden layer, and the threshold value of the hidden layer, the output value of the hidden layer is calculated:

[0009] i is an integer from 1 to 3; j is an integer from 1 to 7;

[0010] In the formula, H j is the output value of the hidden layer; w ij is the weight value from the input layer to the hidden layer; x i is the variable input to the input layer; a j is the threshold value input to the hidden layer; i is the number of input layer nodes; j is the number of hidden layer nodes; f is the hidden layer excitation function, and its expression is:

[0011]

[0012] According to the output value of the hidden layer, the connection weight value between the hidden layer and the output layer, and the threshold value of the output layer, the neural network prediction output result Y' is calculated:

[0013]

[0014] In the formula, H j is the output value of the hidden layer; w jk is the weight value from the hidden layer to the output layer; b k is the threshold value input to the output layer; j is the number of hidden layer nodes; l is the number of hidden layers; k is the number of output layer nodes. Among them, the experimental data obtained by S1 is taken as the training set, and the training set is the sample set for the neural network to learn. Through the calculation of the training set, the law and quantity relationship between the input data and the output data are basically mastered, and then the neural network prediction output result Y' corresponding to different input data is obtained.

[0015] According to the network prediction output result Y i ' and the expected output Yi , calculate the variance E(m):

[0016]

[0017] In the formula: Y i — the expected network output vector; Y i — the actual network output vector; P — the number of samples of the experiment (i.e. the number of experimental groups designed in step S1); m — the vector composed of network weights and thresholds; e i (m) — the error;

[0018] Let m k represent the vector composed of weights and thresholds at the kth iteration, and the vector m k +1 , m k+1 = m k + Δm;

[0019] In the Levenberg-Marquardt method, the weight increment calculation formula is as follows:

[0020] Δw = [J T (w)J(w) + μI] -1 J T (w)e(w)

[0021] In the formula: I — unit matrix; μ — user-defined learning rate; J(w) — Jacobian matrix, i.e.

[0022]

[0023] The threshold calculation method is the same as the weight, i.e. using the above weight increment calculation formula;

[0024] The updated weights and thresholds are continuously calculated according to the above calculation process until the set number of training times is reached; the L-M algorithm can make the mean square error rapidly decrease to 10 -25 in a very short time, greatly reducing the calculation time and quickly obtaining the prediction value.

[0025] Through the above process, a model based on neural network prediction is constructed.

[0026] S3: The experimental data about the vector composed of the three factors and the molecular weight of polylactic acid obtained in step S1 is taken as the training set, and the training set is the sample set for the model based on neural network prediction constructed in step S2 to learn. Through calculation of the training set, the corresponding relationship between the input variable data about the three factors and the output data about the molecular weight of polylactic acid is obtained, and the trained model based on neural network prediction is obtained;

[0027] The experimental data of the vector composed of the three factors and the polylactic acid molecular weight obtained in step S1 is taken as a test set, the three factors in the vector of the test set are taken as input variables, and the trained model is inputted into the trained model to obtain a network prediction output result Y i , which is prediction data; the experimental data obtained in step S1 is taken as expected output Y i , which is actual output data; thus, the coincidence degree of the prediction data and the actual output data is compared by using the data of the test set, so that the effectiveness of the prediction model is judged.

[0028] Further, the reaction temperature x1 in step S1 ranges from 140 to 180 DEG C; the reaction time x2 ranges from 8 to 12 hours; the catalyst addition amount x3 is the ratio of the mass of the catalyst to the mass of the lactic acid liquid, and the change range of x3 is 0.6 to 3.0%, and the polymer molecular weight Y obtained by using the superimposed sampling uniform design method ranges from 1600 to 5600.

[0029] By using the above technology, the present application has the following beneficial effects:

[0030] 1) The present application provides a neural network for predicting polymer reaction conditions, so as to optimize the polylactic acid synthesis reaction conditions. The neural network method provided by the present application for predicting the relationship between the lactic acid polymerization conditions and the polylactic acid molecular weight has the characteristics of high fitting precision and high result reliability, and is greatly helpful for the prediction and optimization of the conditions for preparing polylactic acid with different molecular weights.

[0031] 2) The neural network provided by the present application, as a tool for predicting polymer reaction conditions and molecular weight, has a wide application prospect, and its application range is not limited. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Fig. 1 is a neural network structure diagram for predicting the polylactic acid molecular weight of the present application;

[0033] Figure 2a Fig. 2 is a regression curve diagram of training fitting in Example 2 of the present application;

[0034] Figure 2b Fig. 3 is a test regression curve diagram obtained in Example 2 of the present application;

[0035] Figure 2c Fig. 4 is a total regression curve diagram obtained in Example 2 of the present application. DETAILED DESCRIPTION

[0036] The present application is further described below in combination with specific embodiments, but the protection scope of the present application is not limited to this:

[0037] Example 1

[0038] The zeolite catalyst (zeolite molecular sieve (framework Si / Al molar ratio 50-100), pore size about 0.7-1 nm) is put into the lactic acid liquid for temperature rising reaction, and the related experimental ranges of three factors of reaction temperature (x1), reaction time (x2) and mass fraction of added catalyst (x3, x3 is the ratio of catalyst mass to lactic acid liquid mass) are divided into five levels, the experiment design of three factors and five levels is carried out, and according to the superimposed sampling uniform design method, the reaction conditions of lactic acid polymerization and the molecular weight of the corresponding polylactic acid are shown in Table 1.

[0039] Table 1 3-factor 5-level superimposed sampling uniform design and corresponding polylactic acid molecular weight

[0040]

[0041] Example 2

[0042] A lactic acid polymerization process based on neural network prediction, specifically comprising the following steps:

[0043] S1: According to the data in Table 1 obtained in Example 1, the experimental data about the vector composed of three factors (x1, x2 and x3) and the molecular weight of polylactic acid is obtained;

[0044] S2: A three-layer neural network structure is constructed, the input layer contains 3 nodes corresponding to 3 influencing factors, i.e. reaction temperature x1, reaction time x2 and catalyst addition amount x3; the number of nodes in the hidden layer is determined according to Kolmogorov theorem, the number of nodes in the input layer is n, then the number of nodes in the hidden layer is 2n+1, i.e. 7; the output layer contains one node, i.e. the molecular weight Y of polylactic acid; according to the input variables x1, x2 and x3, the connection weight value between the input layer and the hidden layer, and the threshold value of the hidden layer, the output value of the hidden layer is calculated:

[0045] i is an integer from 1 to 3; j is an integer from 1 to 7;

[0046] In the formula, H j is the output value of the hidden layer; w ij is the weight value from the input layer to the hidden layer; x i is the variable input to the input layer; a j is the threshold value input to the hidden layer; i is the number of input layer nodes; j is the number of hidden layer nodes; f is the hidden layer excitation function, and its expression is:

[0047]

[0048] According to the output value of the hidden layer, the connection weight value between the hidden layer and the output layer, and the threshold value of the output layer, the neural network prediction output result Y' is calculated:

[0049]

[0050] where H j is the output value of the hidden layer; w jk is the weight value of the output layer; b k is the threshold value of the output layer; j is the number of hidden layer nodes; l is the number of hidden layers; k is the number of output layer nodes;

[0051] According to the network prediction output result Y i and the expected output Y i , the variance E(m) is calculated:

[0052]

[0053] where: Y i is the expected network output vector; Y i is the actual network output vector; P is the number of experimental samples; m is the vector composed of network weights and thresholds; e i (m) is the error;

[0054] Let m k represent the vector composed of the weights and thresholds of the kth iteration, and the vector m k +1 , m k+1 = m k + Δm;

[0055] In the Levenberg-Marquardt method, the weight increment calculation formula is as follows:

[0056] Δw = [J T (w)J(w) + μI] -1 J T (w)e(w)

[0057] where: I is the unit matrix; μ is the user-defined learning rate; J(w) is the Jacobian matrix, that is:

[0058]

[0059] The threshold calculation method is the same as the weight, that is, the weight increment calculation formula is used above;

[0060] The updated weights and thresholds are calculated according to the above calculation process until the set number of training times is reached;

[0061] Through the above process, the model based on neural network prediction is constructed;

[0062] S3: using the experimental data of the vector composed of the three factors and the molecular weight of polylactic acid obtained in step S1 as a training set, the training set being a sample set for learning the model based on neural network prediction constructed in step S2, obtaining the corresponding relationship between the input variable data of the three factors and the output data of the molecular weight of polylactic acid by calculating the training set, and obtaining the model based on neural network prediction after training;

[0063] and using the experimental data of the vector composed of the three factors and the molecular weight of polylactic acid obtained in step S1 as a test set, inputting the three factors in the vector of the test set as input variables into the model after training to obtain the network prediction output result Y i , which is the prediction data; and using the experimental data obtained in step S1 as the expected output Y i , which is the actual output data; thus, the coincidence degree of the prediction data and the actual output data is compared by using the data of the test set, so as to judge the effectiveness of the prediction model.

[0064] According to the processes of steps S1-S3, the reaction temperature, reaction time, catalyst addition amount and polymer molecular weight of the first seven groups of experiments in Table 1 of Example 1 are trained and fitted by using the neural network, and the corresponding regression curve diagram of the trained and fitted regression curve is as shown in Figure 2a . Figure 2b is a test regression curve of the reaction temperature, reaction time, catalyst addition amount and polymer molecular weight of the last two groups of experiments in Table 1 of Example 1 by using the neural network, Figure 2c is a total regression curve diagram. The linear correlation degrees of the trained regression curve and the test regression curve are both 1, and the linear correlation degree of the total regression curve is 0.9953. In the diagram, the horizontal coordinate is the molecular weight measured by experiment, represented by T. The vertical coordinate is the molecular weight predicted by the neural network, represented by Y. If the output value of the molecular weight predicted by the neural network is equal to the measured molecular weight, Y=T.

[0065] wherein Figure 2a , Y≈1×Target±2e -12 ; Figure 2b , Y≈1×Target+3.3e 2 ; Figure 2c , Y≈0.99×Target+1.1e 2 . Data is the measured molecular weight and the predicted molecular weight obtained under each group of experimental conditions, and fit is the curve fitted according to the Data value.

[0066] According to the above process, the experimental conditions and results are predicted by using the neural network, and four groups of experimental conditions are selected from the predicted conditions for experimental verification. The selected experimental conditions, prediction results and experimental results are shown in Table 2.

[0067] Table 2 Molecular weight determination results

[0068]

Claims

1. A method for a neural network prediction based lactic acid polymerization process, characterized by The catalyst is put into lactic acid liquid to react at a high temperature, reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as control independent variables, and the molecular weight of polylactic acid is used as a dependent variable, a reaction experiment of synthesizing polylactic acid is carried out through a combination scheme of the three control variables, and experimental data of a vector composed of the independent variables and the dependent variable are obtained; The neural network has a three-layer neural network structure, the model algorithm adopted is Levenberg-Marquardt, reaction temperature x1, reaction time x2 and catalyst addition amount x3 are used as input variables, and the molecular weight of polylactic acid is used as an output variable, a model based on neural network prediction is constructed, the model is used to train the above obtained experimental data, the corresponding relationship between the input variables and the output variables is obtained, and the trained model can predict the molecular weight of polylactic acid synthesized by the combination scheme of the three input variables; The process method specifically comprises the following steps: S1: A shadow sampling uniform design method is used to carry out experimental design of three factors, the synthesis conditions of polylactic acid direct polycondensation are determined by controlling reaction temperature, reaction time and catalyst addition amount, the molecular weight of synthesized polylactic acid is measured, and experimental data of a vector composed of the three factors and the molecular weight of polylactic acid are obtained; S2: A three-layer neural network structure is constructed, the input layer contains 3 nodes corresponding to 3 influencing factors, i.e. reaction temperature x1, reaction time x2 and catalyst addition amount x3; the number of nodes in the hidden layer is determined according to the Kolmogorov theorem, the number of nodes in the input layer is n, then the number of nodes in the hidden layer is 2n+1, i.e. 7; the output layer contains one node, i.e. the molecular weight Y of polylactic acid; the hidden layer output value is calculated according to the input variables x1, x2 and x3, the connection weight value between the input layer and the hidden layer, and the threshold value of the hidden layer: i = an integer from 1 to 3; j = an integer from 1 to 7; where H j is the output value of the hidden layer; w ij is the weight value from the input layer to the hidden layer; x i is the variable input to the input layer; a j is the threshold value input to the hidden layer; i is the number of nodes in the input layer; j is the number of nodes in the hidden layer; and f is the activation function of the hidden layer, which is expressed as: The neural network prediction output result Y' is calculated according to the hidden layer output value, the connection weight value between the hidden layer and the output layer, and the threshold value of the output layer: where H j is the output value of the hidden layer; w jk is the weight value of the hidden layer output to the output layer; b k is the threshold value input to the output layer; j is the number of hidden layer nodes; l is the number of hidden layers; and k is the number of output layer nodes. According to the network prediction output result Y i and the expected output Y i , the variance E(m) is calculated: where: Y i — desired network output vector; Y i — actual network output vector; P— number of samples in the experiment; m— vector of network weights and thresholds; e i (m)— error; Let m k denote the vector of weights and thresholds at the kth iteration, and m k+1 denote the vector of new weights and thresholds k +1 = m k + Δm; In the Levenberg-Marquardt method, the weight increment calculation formula is as follows: Δw = [J T (w)J(w) + μI -1 J T (w)e(w) In the formula, I is a unit matrix, μ is a user-defined learning rate, and J(w) is a Jacobian matrix, i.e. The threshold calculation method is the same as the weight, i.e. the weight increment calculation formula is used; The updated weight and threshold are continuously calculated according to the above calculation process until the set training times are reached; Through the above process, a model based on neural network prediction is constructed; S3: The experimental data of the vector composed of the three factors and the molecular weight of polylactic acid obtained in step S1 are used as a training set, the training set is a sample set for learning of the model based on neural network prediction constructed in step S2, the corresponding relationship between the input variable data of the three factors and the output data of the molecular weight of polylactic acid is obtained through calculation of the training set, and the trained model based on neural network prediction is obtained; The experimental data of the vector composed of the three factors and the polylactic acid molecular weight obtained in step S1 is taken as a test set, the three factors in the vector of the test set are taken as input variables, and the trained model is input into the trained model to obtain a network prediction output result Y i , which is prediction data; and the experimental data obtained in step S1 is taken as expected output Y i , which is actual output data; thus, the coincidence degree of the prediction data and the actual output data is compared by using the data of the test set, so that the effectiveness of the prediction model is judged.

2. The method of claim 1, wherein the neural network is trained using a plurality of training data sets, each training data set comprising a plurality of input data and a corresponding output data, wherein the input data comprises a plurality of process parameters and a plurality of polymer properties, and wherein the output data comprises a plurality of polymer properties. The reaction temperature x1 is in the range of 140-180℃; the reaction time x2 is in the range of 8-12h; the catalyst addition amount x3 is the ratio of the catalyst mass to the lactic acid liquid mass, and x3 is in the range of 0.6-3.0%; and the polymer molecular weight Y obtained by using the superimposed sampling uniform design method is in the range of 1600-5600.

Images