Ternary mixture flash point prediction method based on random forest

Abstract

The invention discloses a ternary mixture flash point prediction method based on a random forest. The method comprises the following steps: determining a flash point value of a ternary mixture experimental sample and establishing a data set; carrying out molecular descriptor calculation on the ternary mixture experiment sample of the data set; dividing the data set into a training set and a test set; screening the molecular descriptors of the training set to obtain the molecular descriptors needing to be reserved; selecting four hyper-parameters of a random forest regression algorithm; training based on the training set and constructing a flash point prediction model; and evaluating and verifying the flash point prediction model. According to the method, the flash point value of the ternary mixture can be efficiently and accurately predicted, necessary safety data in industrial production are provided, the prediction precision is remarkably improved, and safety accidents are avoided.

Description

Technical Field

[0001] The invention relates to a flash point prediction method for a ternary mixture based on random forest. Background Art

[0002] In the chemical industry and engineering, it is vital to ensure that flammable liquids are used safely. Flash point refers to the lowest temperature above which a flammable mixture of vapor and air can form under specific test conditions, and is a key property of liquid flammability hazards. Generally speaking, the lower the flash point, the more volatile the substance, and the greater the risk of fire. In order to prevent fire or explosion in the safety process, it is crucial to have prior knowledge of the flash point of a substance, as it is the main attribute for assessing its flammability hazard.

[0003] Currently, flash point data for pure compounds can be obtained from various sources, such as Lange's Handbook of Chemistry and the DIPPR database of the American Institute of Chemical Engineers. However, in industrial applications, mixtures are more common, and the proportion of each component in the mixture changes when responding to different needs. Although experimental measurement is still the most direct and effective method to determine the flash point of flammable liquids, it is very difficult to perform experimental measurements due to the diversity of mixture products in industrial production, and even the presence of toxicity or radioactivity. Therefore, there is an urgent need to develop safe and reliable methods to evaluate the flash point of mixtures involved in industrial production.

[0004] Early studies mainly used empirical correlation methods to estimate the flash point of substances based on factors such as boiling point, density, and vapor pressure. Although this type of method is simple to calculate, its estimated value often has significant deviations compared with the experimental value. In recent years, a large number of studies have emerged using quantitative structure-property relationship (QSPR) to predict the flash point or other physical and chemical properties of mixtures, showing its good application prospects.

[0005] Currently, QSPR technology mainly focuses on predicting the flash points of pure components or binary mixtures, and only a few studies focus on predicting the flash points of ternary mixtures.

[0006] Predicting the flash point of a ternary mixture is more complex than predicting the flash point of a pure compound or a binary mixture, and appropriate rules are needed to calculate the descriptors of a multicomponent mixture and adopt a high-performance modeling approach. The calculation rules of the multicomponent mixture descriptors need to consider multiple factors rather than simply adding the molecular descriptor values ​​of each component. In addition to the traditional multivariate linear regression (MLR) method for building QSPR models, more and more machine learning techniques have been introduced into QSPR model construction, and optimization algorithms are used to improve their performance. In the field of machine learning, ensemble learning methods have shown superior performance in classification and regression compared with traditional methods. Among these methods, random forest, as an ensemble learning method, has been widely used to handle complex classification or regression tasks. However, there is no report in the literature on the application of random forest to predict the flash point of mixtures. Summary of the invention

[0007] The purpose of the present invention is to address the shortcomings of the current experimental determination method for estimating the flash point of a mixture, such as high cost and danger, as well as the problem of poor accuracy of the empirical method for estimating the flash point of a mixture, and thus propose a method for high-precision prediction of the flash point of a ternary mixture from the perspective of molecular structure based on random forest technology.

[0008] The technical solutions adopted in the present invention are:

[0009] A flash point prediction method for a ternary mixture based on random forest comprises the following steps:

[0010] S1: Determine the flash point value of the ternary mixture experimental sample and establish a data set;

[0011] S2: Molecular descriptor calculation for the ternary mixture experimental samples of the dataset;

[0012] S3: Divide the data set into training set and test set;

[0013] S4: Screening the molecular descriptors of the training set to obtain the molecular descriptors that need to be retained;

[0014] The screening includes two stages: pre-screening and automatic screening. The pre-screening stage is to remove descriptors whose descriptor feature variance is 0 or close to 0 in the training set; the automatic screening stage is to use a combination of genetic algorithm and random forest regression algorithm to determine the descriptor features that are finally retained;

[0015] S5: Select four hyperparameters of the random forest regression algorithm, which correspond to the number of trees, the maximum depth of the tree, the minimum number of samples of internal nodes, and the minimum number of samples of leaf nodes, determine the optimal hyperparameter combination, and use the mean square error of cross-validation on the training set as the evaluation criterion for the quality of the hyperparameter combination;

[0016] S6: Perform training based on the training set in step S5 and build a flash point prediction model;

[0017] S7: Evaluate and validate the flash point prediction model.

[0018] Furthermore, in S1, the ternary mixture experimental sample is formed by mixing different pure components in different proportions; the flash point value of each experimental sample is measured under closed cup conditions using a SYD-261 flash point tester.

[0019] Furthermore, in S2, the calculation of molecular descriptors includes the calculation of molecular descriptors of pure components and the calculation of molecular descriptors of mixtures.

[0020] Further, the pure component molecular descriptor calculation is to use Dragon software to calculate the molecular descriptors (d) of each component for the mixture;

[0021] The mixture molecular descriptor calculation is to determine the mixture molecular descriptor (D) of the ternary mixture using a logarithmic calculation rule. m ), the calculation rules are:

[0022] Furthermore, in S3, the division ratio of the training set and the test set is 4:1.

[0023] Furthermore, in S4, the random forest regression algorithm includes the following steps:

[0024] 41) Randomly obtain N samples from the training set with replacement;

[0025] 42) Randomly select K features from the descriptor features;

[0026] 43) Repeating steps 41) and 42) to obtain a large number of regression trees;

[0027] 44) The average of all regression tree output results is used as the output value of the random forest regression algorithm.

[0028] Furthermore, the regression tree construction process is:

[0029] (1) Assume that the training set D = {(x1, y1), (x2, y2), ..., (x N ,y N )} contains N samples, and the feature vector corresponding to the sample is:

[0030] (2) After sorting all feature vectors by the jth feature of the feature vector, the value s of the jth feature is used as the split point to split the sorted feature vectors into two regions, corresponding to:

[0031] R1(j,s)={x|x(j) ≤s}

[0032] R2(j,s)={x|x (j) >s}

[0033] By solving the formula: Get the best values ​​of j and s;

[0034] Where c1 and c2 are the fixed output values ​​of the corresponding regions after segmentation;

[0035] Since the optimal output value is the average value of the corresponding regional target value, the above formula can be expressed as:

[0036]

[0037] in,

[0038] (3) Select the best j and s to divide the feature space into two regions: R1(j, s) and R2(j, s);

[0039] (4) Calculate the output value of each area: Wherein, m = 1, 2;

[0040] (5) Repeat steps (3) and (4) until the termination condition is reached;

[0041] (6) The feature space of the final training set is divided into M regions {R1, R2, ..., R M}, the regression tree's output

[0042] The output function is expressed as:

[0043]

[0044] in,

[0045]

[0046] Furthermore, in S4, the process of combining the genetic algorithm with the random forest regression algorithm includes:

[0047] (1) Randomly initialize the initial population of N chromosomes of the genetic algorithm;

[0048] (2) Take values ​​for each chromosome and extract the molecular descriptor values ​​in the training set as the input of the random forest regression algorithm; use the cross-validated mean square error as the fitness of each population;

[0049] (3) Perform selection, crossover, and mutation operations in sequence until the stopping condition is met;

[0050] (4) Output the chromosome with the best fitness, thereby obtaining the best feature subset of the molecular descriptor.

[0051] Furthermore, in S5, a grid search method is used to determine the optimal hyperparameter combination.

[0052] Further, in S7, the flash point prediction model is evaluated and verified including: fitting ability, stability and prediction ability.

[0053] The present invention has the following beneficial effects:

[0054] 1) The present invention can efficiently and accurately predict the flash point value of a flammable liquid mixture based only on the composition ratio of the flammable liquid mixture, provide necessary basic data for the production, transportation, storage and use of the flammable liquid mixture in industrial production, and avoid the occurrence of safety accidents such as fire and explosion.

[0055] 2) The present invention incorporates random forests into multiple stages of model construction, which greatly improves the accuracy of flash point prediction of flammable liquid mixtures compared to traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the present invention.

[0057] Figure 2 A comparison chart of the predicted and experimental values.

[0058] Figure 3 A histogram of the prediction errors.

[0059] Figure 4 R is obtained after randomly shuffling the Y value 2 And scatter plot of RMSE indicator.

[0060] Figure 5 Scatter plot of residuals between predicted and experimental values. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with the accompanying drawings.

[0062] like Figure 1 As shown in the flow chart, the present invention provides a flash point prediction method for a ternary mixture based on random forests. Only the molecular structure information of the ternary mixture solution is required to predict its flash point. The specific steps are divided into the following seven steps:

[0063] Step S1: Determination of flash point of a ternary mixture experimental sample set.

[0064] The ternary mixture samples were mixed by different pure components in different proportions, and the flash point value of each sample was measured under closed cup conditions using a SYD-261 flash point tester.

[0065] Step S2: Calculate molecular descriptors for the ternary mixture experimental samples in the data set.

[0066] The calculation of molecular descriptors includes two stages: calculation of pure component molecular descriptors and calculation of mixture molecular descriptors. Specifically:

[0067] (1) In the calculation stage of pure component molecular descriptors, DRAGON software is used to calculate the molecular descriptors of each component for the mixture (d); DRAGON is a professional software widely used for calculating molecular descriptors. It can obtain the descriptor values ​​of various molecules based only on their structural information without the need for experimental data.

[0068] (2) Mixture molecular descriptor calculation stage: The mixture molecular descriptor (D m ), the calculation rules are:

[0069] Step S3: Divide the data set into a training set and a test set;

[0070] In the division of training set and test set, the training set is used to build the flash point prediction model, and the test set is used to evaluate and verify the obtained model. The division adopts random division, and the ratio of training set to test set is 4 to 1.

[0071] Step S4: Screening of molecular descriptors: The screening of molecular descriptors is divided into two stages: pre-screening and automatic screening.

[0072] In the pre-screening stage, the descriptors with a variance of 0 or close to 0 in the training set are first removed, and then the number of descriptor features is further reduced by removing one of the highly correlated descriptor features. The judgment criterion is that the correlation is greater than or equal to 0.95.

[0073] In the automatic screening stage, a method combining genetic algorithm and random forest regression algorithm is used to determine the descriptor features that are finally retained.

[0074] The specific steps of the random forest regression algorithm are:

[0075] 41) Randomly obtain N samples from the training set with replacement;

[0076] 42) Randomly select K features from the descriptor features;

[0077] 43) Repeating steps 41) and 42) to obtain a large number of regression trees;

[0078] 44) The average of all regression tree output results is used as the output value of the random forest regression algorithm.

[0079] The principle of regression tree is as follows:

[0080] Assume that there is a training set D = {(x1, y1), (x2, y2), ..., (x N ,y N )} contains N samples, and the feature vector corresponding to the sample is:

[0081] The regression tree recursively splits the feature space to reach the termination condition, resulting in a corresponding binary tree structure. In this binary tree structure, each node corresponds to a specific part of the training set feature space. The regression tree uses a heuristic method to split the feature space, that is, checking all values ​​of each feature and selecting the split point with the minimum square error.

[0082] Assume that the jth feature x (j) The value s is selected as the segmentation point, then the feature space is segmented into R1(j, s) = {x|x (j) ≤s} and R2(j,s)={x|x (j) >s} two areas.

[0083] By solving the formula:

[0084]

[0085] Get the best values ​​of j and s.

[0086] Where c1 and c2 are fixed output values ​​of the corresponding regions after segmentation.

[0087] Since the optimal output value is the average value of the corresponding regional target value, the above formula can be expressed as:

[0088]

[0089] in,

[0090] Based on the above formula, the process of building a regression tree is as follows:

[0091] 1) Select the best j and s to divide the feature space into two regions: R1(j, s) and R2(j, s);

[0092] 2) Calculate the output value of each area: Wherein, m = 1, 2;

[0093] 3) Repeat 1) and 2) in the sub-region until the termination condition is reached;

[0094] 4) Finally, the feature space of the training set is divided into M regions {R1, R2, ..., R M}.

[0095] The output function of the regression tree is expressed as:

[0096]

[0097] in,

[0098] The specific details of the method combining genetic algorithm and random forest regression algorithm are as follows:

[0099] In the genetic algorithm, the length of each chromosome is the same as the number of molecular descriptors retained after pre-screening, and binary coding is used, with 1 representing the selected feature and 0 representing the selected feature. The solution steps are as follows:

[0100] (1) Randomly initialize an initial population of N chromosomes;

[0101] (2) Individual evaluation: Take values ​​for each chromosome and extract the molecular descriptor values ​​in the training set as the input of the random forest regression algorithm; use the cross-validated mean square error (MSE) as the fitness of each population;

[0102] (3) Perform selection, crossover, and mutation operations in sequence until the stopping condition is met;

[0103] (4) Output the chromosome with the best fitness, thereby obtaining the best feature subset of the molecular descriptor.

[0104] Step S5: The selection of hyperparameters of the random forest regression algorithm involves the selection of four important hyperparameters in the random forest: the number of trees, the maximum depth of the tree, the minimum number of samples of internal nodes, and the minimum number of samples of leaf nodes.

[0105] The above four hyperparameters are all discrete integer values. The grid search method is used to determine the optimal hyperparameter combination, and the mean square error of cross-validation on the training set is used as the evaluation criterion for the quality of the hyperparameter combination.

[0106] Step S6: Construction of a flash point prediction model: The flash point prediction model is constructed based on the training on the training set after the above steps, thereby constructing a flash point prediction model.

[0107] Step S7: Evaluation and verification of the flash point prediction model. The evaluation and verification of the model is a very important step, and the specific method is as follows:

[0108] (1) The fitting ability of the model is measured by using the negative correlation coefficient (R 2 ), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE);

[0109] (2) Model stability, using leave-one-out (LOO) interactive validation and residual plot analysis;

[0110] (3) The predictive ability of the model is measured by the cross-validation coefficient between the target value and the predicted value of the test set samples. The calculation formula is:

[0111]

[0112] In the formula, y i and represent the target value and the predicted value respectively. is the average value of the target values ​​of the training set samples.

[0113] Embodiment 1:

[0114] Taking the ternary miscible aqueous solution as an example, there are seven organic pure substances involved, namely n-butanol, ethanol, propionic acid, n-pentanol, methanol, isopropanol and isobutanol. By selecting two of the seven substances and mixing them with water, a total of six types of mixtures are produced, as shown in Table 1.

[0115] Category No. Component 1 Component 2 Component 3 1 n-Butanol Ethanol water 2 n-Butanol Propionic acid water 3 n-Butanol n-Amyl alcohol water 4 Isobutanol Propionic acid water 5 Methanol Ethanol water 6 Methanol Isopropyl alcohol water

[0116] Table 1

[0117] By adjusting the proportions of the components in Table 1, a total of 341 mixture samples were generated. The flash point values ​​of these mixtures were then tested using a SYD-261 flash point tester under closed cup conditions, and the flash point values ​​of the samples ranged from 283.15K to 345.65K. In the subsequent DRAGON, a total of 1481 molecular descriptors were obtained by calculating the structural parameters of eight substances. After pre-screening, the remaining 284 descriptors were retained, and then automatically screened using a combination of genetic algorithm and random forest regression algorithm. This process set the upper limit of the feature subset to 10 and used 10-fold cross validation. Finally, the 6 most promising descriptors were retained, as shown in Table 2.

[0118] Descriptors definition Descriptor Type RDF030v Radial Distribution Function-030 / weighted by van der Waals volume RDF Descriptor Mor12u Signal 12 / Unweighted 3D-MoRSE Descriptor Mor32m Signal 32 / Weighted by Quality 3D-MoRSE Descriptor Mor09p Signal 09 / Weighted by polarizability 3D-MoRSE Descriptor Mor15p Signal 15 / Weighted by polarizability 3D-MoRSE Descriptor R5u+ R maximum autocorrelation at lag 5 / unweighted GETAWAY Descriptor

[0119] Table 2

[0120] The grid search value list of four important hyperparameters in random forest (number of trees, maximum tree depth, minimum number of samples of internal nodes and minimum number of samples of leaf nodes) is shown in Table 3.

[0121] Parameter name Value list Number of trees 50,60,70,80,90,100 Maximum depth of the tree 5,7,9,11,13,15,17,19 Minimum number of samples for internal nodes 2,4,6,8 Minimum number of samples for leaf nodes 1,2,3,4

[0122] Table 3

[0123] The final hyperparameter combination is determined as: number of numbers 90, maximum tree depth 15, minimum number of internal node samples 2, and minimum number of leaf node samples 1. Subsequently, based on the parameter settings, training is performed on the training set, and the final prediction model is obtained.

[0124] Table 4 lists the evaluation parameters of the model. On the test set, the multiple correlation coefficient is 0.9966, which shows that the prediction model has a strong fitting ability. On the external test set, the result of the interactive validation based on the leave-one-out method is 0.9717, which is slightly lower than the result on the training set, but still shows high accuracy. Further comparison of the RMSE, MAE, and MAPE on the training set and the test set shows that these three indicators have increased significantly. However, these indicators are still far below the acceptable error threshold. Specifically, the allowable error range of MAE is ±10K, while the MAE value of the model is only 2.1961K.

[0125] Figure 2 The figure is a comparison chart of the predicted values ​​and experimental values. The predicted values ​​and experimental values ​​of most samples in the training set and test set are consistent with each other.

[0126] In order to more clearly demonstrate the prediction effect of the model, Figure 3 Provides a histogram of the prediction errors over all samples. Figure 3 It can be seen that the percentage errors of the vast majority of the 341 samples are within 1.5%, and even the prediction errors of 272 samples are within 0.5%. These results confirm the effectiveness of the model.

[0127]

[0128] Table 4

[0129] To verify the absence of chance correlation, a total of 500 y-scrambled tests were performed. Figure 4 The scatter plots of R^2 and RMSE metrics obtained after 500 random shuffles of the Y values ​​are depicted. It is evident from the figure that the proposed model on the random dataset did not show satisfactory R^2 and RMSE values, thus confirming that its superior performance was not due to chance correlation. In order to assess the potential systematic errors in the model development,

[0130] Figure 5 A scatter plot of the residuals between the model predictions and the experimental values ​​was plotted, and it can be seen that most of the data points are randomly distributed around the zero baseline, indicating that there is no systematic error in the model development process.

[0131] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A flash point prediction method for ternary mixtures based on random forests, characterized in that: The steps include: S1: Determine the flash point value of the ternary mixture experimental sample and establish a data set; S2: Molecular descriptor calculation for the ternary mixture experimental samples of the dataset; S3: Divide the data set into training set and test set; S4: Screening the molecular descriptors of the training set to obtain the molecular descriptors that need to be retained; The screening includes two stages: pre-screening and automatic screening, wherein the pre-screening stage is to remove descriptors whose descriptor feature variance in the training set is 0 or close to 0; The automatic screening stage uses a combination of genetic algorithm and random forest regression algorithm to determine the descriptor features that are ultimately retained; S5: Select four hyperparameters of the random forest regression algorithm, which correspond to the number of trees, the maximum depth of the tree, the minimum number of samples of internal nodes, and the minimum number of samples of leaf nodes, determine the optimal hyperparameter combination, and use the mean square error of cross-validation on the training set as the evaluation criterion for the quality of the hyperparameter combination; S6: Perform training based on the training set in step S5 and build a flash point prediction model; S7: Evaluate and validate the flash point prediction model.

2. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S1, the ternary mixture experimental sample is formed by mixing different pure components in different proportions; the flash point value of each experimental sample is measured under closed cup conditions using a SYD-261 flash point tester.

3. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S2, the molecular descriptor calculation includes the calculation of pure component molecular descriptors and the calculation of mixture molecular descriptors.

4. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 3, characterized in that: The pure component molecular descriptor calculation is to use Dragon software to calculate the molecular descriptors of each component for the mixture (d); The mixture molecular descriptor calculation is to determine the mixture molecular descriptor (D) of the ternary mixture using a logarithmic calculation rule. m ), the calculation rules are:

5. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S3, the ratio of training set to test set is 4:

1.

6. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S4, the random forest regression algorithm includes the following steps: 41) Randomly obtain N samples from the training set with replacement; 42) Randomly select K features from the descriptor features; 43) Repeating steps 41) and 42) to obtain a large number of regression trees; 44) The average of all regression tree output results is used as the output value of the random forest regression algorithm.

7. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 6, characterized in that: The regression tree construction process is: (1) Assume that the training set D = {(x1, y1), (x2, y2), ..., (x N ,y N )} contains N samples, and the feature vector corresponding to the sample is: (2) After sorting all feature vectors by the jth feature of the feature vector, the value s of the jth feature is used as the split point to split the sorted feature vectors into two regions, corresponding to: R1(j,s)={x|x (j) ≤s} R2(j,s)={x|x (j) >s} By solving the formula: Get the best values ​​of j and s; Where c1 and c2 are the fixed output values ​​of the corresponding regions after segmentation; Since the optimal output value is the average value of the corresponding regional target value, the above formula can be expressed as: in, (3) Select the best j and s to divide the feature space into two regions: R1(j, s) and R2(j, s); (4) Calculate the output value of each area: Wherein, m = 1, 2; (5) Repeat steps (3) and (4) until the termination condition is reached; (6) The feature space of the final training set is divided into M regions {R1, R2, ..., R M }, the output function of the regression tree is expressed as: in, 8. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S4, the process of combining the genetic algorithm with the random forest regression algorithm includes: (1) Randomly initialize the initial population of N chromosomes of the genetic algorithm; (2) Take values ​​for each chromosome and extract the molecular descriptor values ​​in the training set as the input of the random forest regression algorithm; use the cross-validated mean square error as the fitness of each population; (3) Perform selection, crossover, and mutation operations in sequence until the stopping condition is met; (4) Output the chromosome with the best fitness, thereby obtaining the best feature subset of the molecular descriptor.

9. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S5, the grid search method is used to determine the optimal hyperparameter combination.

10. The flash point prediction method for ternary mixtures based on random forests as claimed in claim 1, characterized in that: In S7, the flash point prediction model was evaluated and validated including: fitting ability, stability and prediction ability.

Images