Modeling and predicting method based on machine learning azeotrope system judgment
By constructing an azeotropic system determination model using machine learning methods, the problems of long prediction time and high cost of azeotropic systems are solved, achieving fast and accurate prediction of azeotropic systems, simplifying feature engineering, and expanding the prediction range.
Patent Information
- Application Number
- CN202511301930.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-12
- Publication Date
- 2025-12-19
AI Technical Summary
Existing technologies are time-consuming, costly, and have limited predictive range in predicting azeotropic systems, while machine learning has shortcomings in feature engineering and model optimization.
We employ methods including data preprocessing, feature construction, neural network model construction and training, hyperparameter optimization, and model evaluation. We establish an azeotropic system determination model through machine learning, utilize physical properties such as polarizability, boiling point, and hydrogen bond acidity to construct mixture characteristics, and optimize model hyperparameters to improve prediction accuracy and generalization ability.
It enables rapid and low-cost prediction of azeotropic systems, improves the prediction accuracy and robustness of the model, adapts to complex nonlinear relationships, simplifies the complexity of manual design, and expands the prediction range.
Smart Images
Figure CN121171394A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of azeotrope system prediction, in particular to a modeling and prediction method for azeotrope system judgment based on machine learning, which is used for predicting whether a binary mixture forms an azeotrope system. BACKGROUND
[0002] Azeotropy is ubiquitous in nature and industrial processes, which refers to the phenomenon that a binary or multi-component liquid mixture with a specific ratio boils at a constant pressure with the same vapor composition as the solution. In fact, the vapor produced at this time is exactly the same as the liquid composition. Azeotrope cannot be separated by conventional distillation or fractional distillation. The prediction of azeotrope system is of great significance in the fields of chemical engineering, materials science, etc. Traditional methods rely on experimental data and theoretical models, which have problems such as long time consumption, high cost, limited prediction range, etc. In recent years, machine learning has shown great potential in molecular property prediction and property modeling, but existing research still has deficiencies in feature engineering, model optimization and evaluation, especially in handling mixture input features and optimizing model hyperparameters. SUMMARY
[0003] The purpose of the present application is to solve the above-mentioned problems in the prior art, and to provide a modeling and prediction method for azeotrope system judgment based on machine learning, which realizes the rapid prediction of the azeotropic properties of binary mixtures through data preprocessing, feature construction, model training and optimization, reduces the judgment cost, expands the prediction range, and meets the actual needs of industrial scenarios.
[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:
[0005] A modeling and prediction method for azeotrope system judgment based on machine learning, comprising the following steps:
[0006] 1) Data preprocessing: processing the collected azeotrope system experimental sample data, including deleting samples that do not meet the quality requirements and converting sample labels, to obtain preprocessed data;
[0007] 2) Feature construction: constructing mixture input features based on the physical properties of each pure component in the binary mixture;
[0008] 3) Neural network model construction and training: adopting a neural network model connected in turn by input layer-hidden layer-output layer, and configuring an optimizer to train the model;
[0009] 4) Hyperparameter optimization: optimizing the model hyperparameters by using the grid search method to determine the optimal hyperparameter combination;
[0010] 5) Model evaluation: Divide the dataset and use preset evaluation metrics to evaluate the performance of the trained model, and output the azeotropic determination result of the binary mixture.
[0011] Step 1) The data preprocessing specifically includes:
[0012] 1.1) Data cleaning: Remove samples containing missing or duplicate values to ensure data uniqueness and integrity;
[0013] 1.2) Sample label processing: Convert the category labels of "azeotropic" and "non-azeotropic" in the samples into binary labels;
[0014] 1.3) Standardization: The input features constructed subsequently are standardized using the formula (X-μ) / σ, where X is the original value of the feature, μ is the mean of the feature, and σ is the standard deviation of the feature, in order to eliminate the impact of feature scale differences on model performance.
[0015] Step 2) The feature construction specifically includes:
[0016] 2.1) Determine the physical properties of each pure component in the binary mixture, including polarizability, boiling point, hydrogen bond acidity, hydrogen bond basicity, topological polar surface area (TPSA), and dipole moment.
[0017] 2.2) For each physical property, calculate the "absolute difference" and "sum value" of the property for the two pure components in the binary mixture. Use the "absolute difference" and "sum value" as input features of the mixture and delete redundant features to simplify the model input.
[0018] Step 3) The construction and training of the neural network model specifically includes:
[0019] 3.1) Model activation function configuration: The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function;
[0020] 3.2) Optimizer and loss function configuration: The Adam optimizer is used to optimize the model, and the cross-entropy loss function is used as the basis for calculating the loss during model training;
[0021] 3.3) Model training: Input the preprocessed feature data into the neural network model, train the model based on the configured optimizer and loss function, and fit the azeotropic determination law of binary mixtures.
[0022] Step 4) The hyperparameter optimization specifically includes:
[0023] 4.1) Determine the types of hyperparameters to be optimized: The hyperparameters to be optimized include the number of hidden layer nodes, Dropout probability, number of iterations, and batch size;
[0024] 4.2) Hyperparameter combination selection: By traversing the preset hyperparameter range through grid search, the model performance under different hyperparameter combinations is compared, and the optimal hyperparameter combination is determined to be: 50 hidden layer nodes, 0.2 Dropout probability, 200 iterations, and 128 batch size.
[0025] Step 5) The model evaluation specifically includes:
[0026] 5.1) Dataset partitioning: The preprocessed sample data is divided into training set, validation set and test set according to a preset ratio. The training set is used for model fitting, the validation set is used for hyperparameter selection, and the test set is used for final performance evaluation.
[0027] 5.2) Evaluation metrics selection: Accuracy, precision, recall, and F1 score are used as evaluation metrics for model performance;
[0028] 5.3) Evaluation results requirements: After verification on the test set, the model's accuracy in determining "non-azeotropic samples" is no less than 92%, and the accuracy in determining "azeotropic samples" is no less than 90%, ensuring that the model has reliable azeotropic determination capabilities.
[0029] Compared with the prior art, the beneficial effects achieved by the technical solution of this invention are:
[0030] 1. This invention proposes a modeling and prediction method for determining azeotropic systems based on machine learning. Through data preprocessing, feature construction, model training, and optimization, it effectively predicts whether a mixture is azeotropic. The model established using machine learning can easily determine whether a mixture of two components has azeotropic characteristics. This method is simple, convenient, and easy to implement, overcoming the cumbersome steps of experimentally determining whether two components form an azeotrope.
[0031] 2. This invention utilizes machine learning methods to establish a predictive model for whether a mixture forms an azeotrope. Preprocessed data on the components' boiling point, polarizability, dipole moment, hydrogen bond acidity, hydrogen bond basicity, and topological polar surface area (TPSA) are used as model input. The output of the established model determines whether the components form an azeotrope. Training and parameter optimization of the predictive model improve its accuracy and generalization ability. Feature construction and standardization during the modeling process effectively handle the mixture's input features, enhancing the model's predictive performance and robustness.
[0032] 3. This invention employs machine learning algorithms that can capture complex nonlinear relationships, adapt to the complex mechanisms of azeotropic system formation, and effectively handle high-dimensional, multi-feature data. Through the algorithm, the contribution of each input feature (such as polarizability, boiling point, hydrogen bond acidity, etc.) to azeotropic formation can be evaluated, helping to understand the physicochemical mechanism. In addition, through reasonable data partitioning (training set, validation set, test set) and hyperparameter optimization (such as the number of hidden layer nodes, activation function, dropout probability), the model can demonstrate good predictive ability on unknown data.
[0033] 4. This invention utilizes machine learning methods to combine feature engineering (such as absolute difference and summation to construct features) to automatically extract mixture features from pure component properties, reducing the complexity of manual design. Attached Figure Description
[0034] Figure 1 Here is a flowchart of the data preprocessing process;
[0035] Figure 2 A schematic diagram of the feature construction method;
[0036] Figure 3 The optimal model network structure diagram;
[0037] Figure 4 The model's learning curve;
[0038] Figure 5 The impact of activation functions on model performance;
[0039] Figure 6 The impact of batch size on model performance;
[0040] Figure 7 The impact of the number of iterations on model performance;
[0041] Figure 8 The impact of Dropout probability on model performance;
[0042] Figure 9 This illustrates the impact of the number of hidden layer nodes on model performance. Detailed Implementation
[0043] To make the technical problems, technical solutions and beneficial effects of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0044] Example 1
[0045] This embodiment presents a modeling and prediction method for determining azeotropic systems based on machine learning, comprising the following steps:
[0046] 1. Data Preprocessing
[0047] See Figure 1 The data preprocessing workflow involves collecting sample data for modeling from literature or physical property databases. Specifically, azeotropic system data, including 132 substances and 1899 samples, was collected from public databases. Samples with missing or duplicate values were removed to ensure data uniqueness and integrity. The "azeotropic" and "non-azeotropic" category labels in the samples were converted to binary labels. The input features for subsequent construction were standardized using the formula (X-μ) / σ, where X is the original value of the feature, μ is the mean of the feature, and σ is the standard deviation of the feature, to eliminate the impact of feature scale differences on model performance.
[0048] 2. Feature Construction
[0049] See Figure 2 The feature construction process mainly includes feature variable selection, feature construction, and feature optimization. Feature variables are mainly selected from the physical properties of each component in the mixture. Physical properties mainly include polarizability, boiling point, hydrogen bond acidity, hydrogen bond basicity, topological polar surface area (TPSA), and dipole moment. For each physical property, the "absolute difference" and "sum value" of this property for the two pure components in the binary mixture are calculated. The "absolute difference" and "sum value" are used as the input features of the mixture, and redundant features are deleted to simplify the model input.
[0050] 3. Neural Network Model Construction and Training
[0051] A neural network prediction model was constructed, consisting of an input layer, hidden layers, and an output layer. The hidden layers used the ReLU activation function, and the output layer used the Softmax activation function. The model was trained using the Adam optimizer and the cross-entropy loss function to fit the azeotropic determination rules of binary mixtures.
[0052] Figure 3 Describe the neural network structure of the prediction model for azeotropic systems, which includes an input layer, a hidden layer, and an output layer.
[0053] Figure 4 The model learning curves generated from 10 random number seeds are presented, indicating that the model did not exhibit significant overfitting during training.
[0054] Figure 5 The description of activation functions refers to the trend of model accuracy as a function of different activation functions. Activation functions include ReLU, Sigmoid, Tanh, ELU, and LeakyReLU, with ReLU being the best performing activation function.
[0055] 4. Hyperparameter optimization
[0056] The types of hyperparameters to be optimized were determined: the number of hidden layer nodes, the Dropout probability, the number of iterations, and the batch size; the hyperparameters were optimized through grid search, and the optimal combination was determined: 50 hidden layer nodes, 0.2 Dropout probability, 200 iterations, and 128 batch size.
[0057] Figure 6 This section describes the trend of model accuracy on the validation set as batch sizes of 32, 64, 128, 256, 512, and 1024 change with the batch size. Batch size refers to the number of samples used in each iteration of model training. A larger batch size can speed up model training, but the model tends to memorize larger batches of data, leading to decreased generalization ability. A smaller batch size requires more iterations to complete training.
[0058] Figure 7 This indicates the performance trend of the model under different iteration numbers when the number of iterations is set to 50-250 with an interval of 10. The correct selection of the number of iterations during model training can ensure that the features of the training set are fully learned and the model parameters are gradually optimized during the training process.
[0059] Figure 8 The description shows the performance trend of the model when the Dropout probability is set to 0.1-0.6 with an interval of 0.1. From this trend, it can be determined that when the Dropout probability is set to 0.2, the model's prediction accuracy reaches its highest level, and the model's robustness is also good.
[0060] Figure 9 The accuracy trends of the model are described when the number of hidden layer nodes is set to 10, 20, 30, 40, 50, and 60, respectively. It can be determined that the average accuracy of the model reaches its highest point and the standard deviation is also the lowest when the number of hidden layer nodes is 50.
[0061] 5. Model Evaluation
[0062] The preprocessed sample data was divided into training, validation, and test sets according to a preset ratio. The training set was used for model fitting, the validation set for hyperparameter selection, and the test set for final performance evaluation. Accuracy, precision, recall, and F1 score were used to evaluate model performance to ensure predictive ability and reliability. The model prediction results are shown in Table 1.
[0063] Table 1
[0064]
[0065] This invention proposes a modeling and prediction method for azeotropic system determination based on machine learning. Through data preprocessing, feature construction, model training, and optimization, it effectively predicts whether a mixture is azeotropic. The model established using machine learning can easily determine whether a mixture of two components has azeotropic characteristics. This method is simple, convenient, and easy to implement, overcoming the cumbersome steps of experimentally determining whether two components form an azeotrope.
Claims
1. A modeling and prediction method for azeotropic system determination based on machine learning, characterized in that, Includes the following steps: 1) Data preprocessing: The collected experimental sample data of azeotropic systems are processed, including deleting samples that do not meet the quality requirements and changing the sample labels to obtain preprocessed data; 2) Feature Construction: Based on the physical properties of each pure component in the binary mixture, construct the input features of the mixture; 3) Neural network model construction and training: A neural network model with input layer-hidden layer-output layer connected in sequence is adopted, and an optimizer is configured to train the model; 4) Hyperparameter optimization: The model hyperparameters are optimized using a grid search method to determine the optimal combination of hyperparameters; 5) Model evaluation: Divide the dataset and use preset evaluation metrics to evaluate the performance of the trained model, and output the azeotropic determination result of the binary mixture.
2. The modeling and prediction method for determining azeotropic systems based on machine learning as described in claim 1, characterized in that, Step 1) The data preprocessing specifically includes: (1.1) Data cleaning: Delete samples containing missing or duplicate values to ensure data uniqueness and integrity; (1.2) Sample label processing: Convert the category labels of "azeotropic" and "non-azeotropic" in the sample into binary labels; (1.3) Standardization: The input features constructed subsequently are standardized using the formula (X-μ) / σ, where X is the original value of the feature, μ is the mean of the feature, and σ is the standard deviation of the feature, in order to eliminate the impact of feature scale differences on model performance.
3. The modeling and prediction method for determining azeotropic systems based on machine learning as described in claim 1, characterized in that, Step 2) The feature construction specifically includes: (2.1) Determine the physical properties of each pure component in the binary mixture, including polarizability, boiling point, hydrogen bond acidity, hydrogen bond basicity, topological polar surface area (TPSA) and dipole moment; (2.2) For each physical property, calculate the "absolute difference" and "sum value" of the property for the two pure components in the binary mixture. Use the "absolute difference" and "sum value" as the input features of the mixture and delete redundant features to simplify the model input.
4. The modeling and prediction method for determining azeotropic systems based on machine learning as described in claim 1, characterized in that, Step 3) The construction and training of the neural network model specifically includes: (3.1) Model activation function configuration: The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function; (3.2) Optimizer and loss function configuration: The Adam optimizer is used to optimize the model, and the cross-entropy loss function is used as the basis for loss calculation during model training; (3.3) Model training: Input the preprocessed feature data into the neural network model, train the model based on the configured optimizer and loss function, and fit the azeotropic determination law of binary mixtures.
5. The modeling and prediction method for determining azeotropic systems based on machine learning as described in claim 1, characterized in that, Step 4) The hyperparameter optimization specifically includes: (4.1) Determine the types of hyperparameters to be optimized: The hyperparameters to be optimized include the number of hidden layer nodes, Dropout probability, number of iterations, and batch size; (4.2) Hyperparameter combination selection: By traversing the preset hyperparameter range through grid search, the model performance under different hyperparameter combinations is compared, and the optimal hyperparameter combination is determined to be: 50 hidden layer nodes, 0.2 Dropout probability, 200 iterations, and 128 batch size.
6. The modeling and prediction method for determining azeotropic systems based on machine learning as described in claim 1, characterized in that, Step 5) The model evaluation specifically includes: (5.1) Data set partitioning: The preprocessed sample data is divided into training set, validation set and test set according to a preset ratio. The training set is used for model fitting, the validation set is used for hyperparameter selection, and the test set is used for final performance evaluation. (5.2) Evaluation index selection: Accuracy, precision, recall and F1 score are used as evaluation indexes for model performance; (5.3) Evaluation results requirements: After verification on the test set, the model's accuracy in determining "non-azeotropic samples" is not less than 92%, and the accuracy in determining "azeotropic samples" is not less than 90%, ensuring that the model has reliable azeotropic determination capabilities.