KR101295861B1 - Multiple Linear Regression―Artificial Neural Network Hybrid Model Predicting Lower Flammability Limit Volume Percent of Organic Compound - Google Patents

Abstract

The present invention consists of up to five elements such as hydrogen (H), carbon (C), nitrogen (N), oxygen (O), and sulfur (S), and is composed of pure molecules consisting of up to 25 atoms except hydrogen. A mathematical model is provided to predict with high accuracy the lower flammability limit volume percent of organic compounds. The above model is based on the lower limit of the volumetric volume percentage as an independent variable for a plurality of organic compounds satisfying the above-mentioned conditions for which the experimental value of the lower volumetric volume limit is known. The best of many multiple linear regression models is obtained by using a genetic algorithm, and then the value of the molecular descriptors included in this model is input to output the lower volume percent of print. It is an example of a quantitative structure-property relationship (QSPR) model, which is a hybrid model of multiple linear regression-artificial neural network that further improves predictive performance by constructing an artificial neural network. Lower bound volume percentage of any molecule that is purely composed of this molecule Prediction give to. As such, the present invention provides a method for predicting the value of the lower limit of the flammability volume value even for a large number of organic compounds in the above-mentioned conditions, in which the experimental value is unknown, thereby saving the cost and time required for the experiment, and research and development in related industries. It produces effects such as facilitating activities.

Description

Multiple Linear Regression-Artificial Neural Network Hybrid Model Predicting Lower Flammability Limit Volume Percent of Organic Compound}

The present invention belongs to a field of physical chemistry called physical property prediction and relates to a method for predicting with high accuracy a lower flammable volume percentage, which is one of several physical properties of an organic compound.

Knowing the exact values of the various properties of an organic compound in detail can be used to determine the feasibility of the use of the substance, to design the synthesis and purification processes, and to establish methods and conditions for storage, transport, use, and disposal. It is an important issue both industrially and academically because it is critical to all decision-making processes throughout the process. The most accurate way of knowing the value of the property of interest of an organic compound of interest is also an experiment, but it is true that it is quite costly and time consuming in many aspects, including the preparation of purified samples and the construction of an environment for accurate measurement. Therefore, it may not be possible. Thus, as an alternative, many researchers have long tried to predict the exact value of various physical properties of organic compounds. As such, property prediction has a long history, and new prediction methods are constantly appearing. Currently, prediction models with different accuracy and scope of application coexist.

Several predictive models have been proposed to date for lower flammable volume percentages, which are the properties of interest of the present invention. The flammability range is divided into a lower flammability limit and an upper flammability limit, which is the lower flammability limit volume percent when the flammable gas is mixed with air, the concentration at the lowest point that generates a gas that can be continuously flammable by the ignition source. It is called.

To date, no group contribution model has been reported for lower flammable volume percentages and one alternative is the quantitative structure-property relationship (QSPR) method. This method basically starts with the assumption that the properties of a compound are a function of the structural properties of the molecule and uses various molecular descriptors that reflect different structural properties. The number of molecular descriptors proposed to date has been thousands, and many kinds of molecular descriptors range from simple ones such as the number of carbons or hydrogens in a molecule to complex ones such as the shape, connection state, and electrochemical properties of molecules. Calculations have been developed for Todeschini R., V. Consonni V., Molecular Descriptors for Chemoinformatics: Second , Revised and Enlarged Edition : Volume I / II , Wiley-VCH, 2009]. The QSPR prediction model is presented in the form of a function that includes some of these molecular descriptors and sometimes, in addition, some of the other physicochemical properties of the compound, which are also functions of structural properties.

들의 선형 결합 함수이며 각 계수

들은 주로 다중선형회귀분석을 통해 실험데이터로부터 결정된다.The most frequently adopted form of such a function is

Linear coupling function of each coefficient

These are mainly determined from experimental data through multiple linear regression analysis.

이 부여되어 있다. 은닉층과 출력층의 각 노드들은 전 단계의 노드들로부터 이러한 연결들을 통해 입력을 받은 뒤 이를 가공하여 출력값을 만드는데 이때 활성화 함수(activation function)라 불리는 함수

를 적용한다. 이러한 인공신경망을 실제로 활용하려면 먼저, 다양한 입력값과 그 입력값에 대응하는 출력값을 함께 묶어 놓은 샘플집합을 이용하여 인공신경망을 훈련시키는 과정이 필요한데 이는 주어진 입력에 대한 인공신경망의 출력과 원하는 출력의 차이가 최소가 되도록 역전파(back propagation) 알고리즘을 사용하여 각 연결의 가중치를 최적화하는 것을 말한다. 이러한 훈련을 거친 인공신경망은 문제해결에 필요한 규칙이나 지식을 따로 제공하지 않아도 학습을 통해서 스스로 일반적인 규칙을 수립하여 미지의 입력에 대해서도 타당성 있는 출력을 내주므로 화합물의 물성예측과 같이 기반 이론이 결여되어 있는 분야에 매우 유용한 수단으로 널리 이용되고 있다.Another way to create a QSPR model is to use an artificial neural network. Artificial neural network technology is one of the long-standing research results of human beings who want to make intelligent machines by modeling human nerve cells with intelligence, and is an information processing technology that has been applied in various fields since the mid 20th century. . 2 shows a typical example of an artificial neural network. As can be seen, the artificial neural network has an input layer for receiving input data, an output layer for producing output data, and a hidden layer located between them, each layer having one or more nodes. ) Each node in the hidden layer is connected to nodes in the input and output layers, and each connection has a quantity called a weight.

Is granted. Each node in the hidden and output layers receives input through these connections from the nodes in the previous stage and then processes it to produce an output, which is called an activation function.

Is applied. In order to actually use such an artificial neural network, first, the neural network is trained using a sample set in which various input values and output values corresponding to the input values are bundled together. This means optimizing the weight of each connection using a back propagation algorithm to minimize the difference. The neural network that has undergone such training does not provide the necessary rules or knowledge to solve the problem, but it establishes general rules through learning and gives valid output for unknown inputs. It is widely used as a very useful means in the field.

The QSPR prediction model for lower flammable volume percentages has been proposed, but not yet widely used. Gharagheizi proposed a genetic algorithm-multilinear regression model using molecular descriptors (Gharagheizi F., Energy & Fuels , 22, 5, (2008)), but the predictive performance was not very good. Later, we applied artificial neural network technique to group contribution method [Gharagheizi F., J. Hazard . Mater . , 170, 595 (2009)], 1057 subjects, and more improved results than before, but 105 pieces were used, and the problems of group contribution methods remain.

Pan and his colleagues proposed a QSPR prediction model for pure hydrocarbons [Pan Y., Jiang J., Ding X., Wang R., Jiang J., AIChE J., 56, 3 (2010)]. Dogs are less diverse and have poor predictive performance.

The technical problem to be achieved by the present invention is more reliable, based on more experimental data, such as hydrogen (H), carbon (C), nitrogen (N), oxygen (O), sulfur (S) within five elements It provides a QSPR model for the lower percent flammable volume of a pure organic compound consisting of molecules of up to 25 atoms excluding hydrogen.

This model is expected to overcome the shortcomings of the models constructed by the group contribution method as mentioned above, and to show wider and more accurate prediction performance. The above-mentioned limitations on the range of organic compounds to which the predictive model can be applied are mainly due to the current state of the art when some of the molecular descriptors used require quantum mechanical calculations. This is due to the fact that for the compounds beyond the stated range, problems arise in terms of accuracy and calculation time. However, even within the above limitations, since there are a great many compounds and industrially important compounds are included in a large amount, it is determined that the present invention can greatly benefit human society.

Humans today depend on a wide variety of organic compounds, including plastics, fibers, rubber, paints, fertilizers, medicines and fuels, and this trend is expected to intensify. According to the American Chemical Society (ACS), as of July 2010, the total number of compounds registered was over 54,000,000. In comparison, even if the physical property value is only one, the number of experimentally known compounds is only tens of thousands. The physical property value of compounds is essential for the better life of mankind, such as the development of new materials and new drugs, the optimal design of chemical plants, the improvement of the productivity of existing facilities, the development and saving of resources, the securing of safety, and the protection of the environment.

However, at present, the number of compounds whose experimental values are known is only a few thousand, and depending on the compounds, the work of obtaining data through experiments may be past due to toxicity, instability, and difficulty of purification. From this point of view, the present invention, which provides a high accuracy of the volume percent lower limit of flammability of a large number of compounds with only information on the molecule, does not only reduce the cost and time required for the experiment, but also estimates the value even when the experiment is impossible. It will not only facilitate the R & D activities of related industries, but also provide the information that can be carried out more smoothly by providing reasonable information to all those who need the value, such as academics and academia. .

1 is a flowchart illustrating a process of building a QSPR model of a lower flammable volume percentage provided by the present invention.
Figure 2 is a view showing the structure of the artificial neural network used in the present invention.
FIG. 3 is a parity diagram of 1298 experimental data of a QSPR model of the lower flammable volume percentage provided by the present invention. FIG.
4 is a histogram diagram of 1298 experimental data of the QSPR model.

Figure 1 is a simplified representation of the process of building a QSPR model for the lower flammable volume percentage.

The first thing to do when building the model is to collect and review the experimental data as specified in step 1. As a result of extensive research on all the literatures and data that can be referred to through various papers, books, internet sites, etc. for the present invention, the data on the lower limit of volume of flammability for the total of 1200 compounds meeting the conditions of the present invention Collected. The collected data were examined in a variety of ways to determine whether they were truly valid values that could be used to build the model.They were not experimental values, errors in data notation, or values for the same compound. If the values are unreliably distant or if the values for the molecular expressions are data about compounds that are difficult to prepare immediately, the data is modified or deleted through a thorough analysis to finally obtain a total of 1190 (10 out of 10) compounds. Data were selected. In addition, when constructing a property prediction model, it was better to classify sample compounds separately into hydrocarbons consisting of carbon and hydrogen and nonhydrocarbons, and to model them separately. In light of this, we decided to classify them into 356 hydrocarbons and 834 non-hydrocarbons, respectively, to establish a model. In the present invention, the 'organic compound' or 'compound' is composed of five elements such as hydrogen (H), carbon (C), nitrogen (N), oxygen (O), and sulfur (S). Refers to a substance consisting of molecules with 25 or fewer atoms excluded.

The next step is to prepare the values of the molecular descriptors for these compounds. The values for various molecular descriptors totaling up to 1978 are computed in batches using a computer from files containing information on the molecules of each compound. In order to calculate the electronic structure of molecules, the Schrodinger equation is usually solved by a pure theory to solve the electron energy. However, in the case of systems with many electrons, Hartley uses an approximation method that ignores electron correlation. Hartree-Fock (HF) method [CC J. Roothan, Rev. Mod. Phys. 23, 69 (1951)]. This approximation introduces a fundamental error in the calculated results and adds a multidimensional theoretical perturbation term to the Post Hartree-Fock method [C. Moller and M. S. Plesset, Phys. Rev. 46, 618 (1934)], to obtain a more accurate solution, but require a relatively large amount of computation. In this way, it is too costly or time-consuming to calculate large molecules.

In addition, the Gaussian method that combines Hartley-Fork and Post Hartley-Fork [L. A. Curtiss, K. Raghavachari, G. W. Trucks, and J. A. Pople, J. Chem. Phys. 94, 7221 (1991); L. A. Curtiss, K. Raghavachari, P. C. Redfern, V. Rassolov, and J. A. Pople, J. Chem. Phys. 109, 7764 (1998) show very little error in energy prediction, but more computation is required because it performs energy calculations for several post-Hartley-Fork methods.

Density functional theory is used to find the ground state using the function of the total energy using the electron density function instead of the wave function with the multidimensional perturbation term to consider the correlation between the electrons of the molecules of many electrons. R. Seeger and JA Pople, J. Chem. Phys. 66, 3045 (1977). The advantage of the density functional theory is that the electron density only needs to be taken into account so that more accurate results can be obtained with comparable calculations to the Hartree-Fock method. The combination of exchange functions and correlation functions for calculating the exchange-correlation energy of the electrons is used to obtain more improved results without increasing the calculation amount.

The computational theories previously attempted to select an optimal quantum mechanical calculation method are the density ranges of the aforementioned Hartley-Fork method, various Post-Hartley-Fork methods, Gaussian (G2, G3) methods, and various combinations of functional functions. Function theory. Among them, one method of density functional theory, which is the best performance calculation time, was selected.

Therefore, in the present invention, the optimization of the molecular structure and the frequency calculation are performed by applying the calculation method of the specified density functional theory using a commercial quantum mechanical calculation program.

In the optimized structure, molecular descriptors represented by various meaningful values reflecting the properties of molecules as well as the above-described physical property information can be obtained. Some molecular descriptors can express the characteristics of two-dimensional structures, while others represent the characteristics of three-dimensional structures. Divided into 24 categories, including detailed presenters in each category. After calculating the molecular descriptor values, we picked out those that were not suitable, that is, the values were the same for all sample compounds and could not be independent variables in the model. This is because it prevents irrelevant molecular expressions from being included in the prediction model, thereby increasing the reliability of the model and reducing the computation time required to find the optimal model by reducing the number of molecular expressions.

In step 4, the sample compounds are divided into two parts: the training set used to explore the prediction model, and the test set used to test the predictive performance of the model determined by the training set. The sample hydrocarbons and the non-hydrocarbons were divided between 5: 5 and 8: 2, preferably in a ratio of 7 to 3, taking care not to distribute similar molecules in only one portion.

이며 이들을 다 조사하는 것은 현실적으로 불가능하다.We then find the best multiple linear regression model based on the training set. The term 'best' is used in the sense of relative meaning that it can be obtained in a relatively short time and has a performance very close to the optimal solution in the absolute sense. The reason for not finding the optimal solution directly is because of the long computation time. For example, when there are 1700 suitable molecular representations out of 1978 molecular representations, you can choose from five different linear regression models. The total number is

It is practically impossible to investigate them all.

In order to obtain useful results within a limited time, the present invention uses a genetic algorithm [Judson, "Genetic Algorithms and Their Uses in Chemistry", Reviews in Computational Chemistry, Lipkowitz & Boyd, Eds., Vol. 10, pp. 1 -73 (VCH Publishers, NY, 1997)]. First, a population of multiple linear regression models is created by randomly drawing a certain number of molecular descriptors from a pool of molecular descriptors. For example, let's say we created a population of 1000 different polylinear regression models that were randomly drawn from 5 of the 1700 suitable molecular descriptors.

Individuals, called chromosomes, are coded by combining the numbers of extracted molecular descriptors. For example, the chromosome of the multiple linear regression model formed by the 45th, 167, 684, 1033, and 1502th molecular descriptors among 1700 molecular descriptors can be expressed as (45, 167, 684, 1033, 1502). Two parent chromosomes are selected from the populations thus generated and crossed over to generate children. In the present invention, the Roulette Wheel method is adopted as a selection method for selecting the parent chromosomes.

) 또는 평균절대오차(average absolute error: AAE)를 활용하였다. 즉 결정계수값이 크거나 평균절대오차값이 작은 것이 선택확률이 높도록 하였다.The Roulette Wheel method is the most commonly used selection algorithm, which allocates a roulette region to the chromosome in proportion to the fitness of each chromosome, and then rotates the roulette to select the chromosome of the corresponding region. Therefore, in this method, the higher the fit, the more likely it is to be selected. The coefficient of determination of the regression model is used to calculate the goodness of fit for each chromosome that determines the probability of selection.

) Or average absolute error (AAE). In other words, the larger the coefficient of determination or the smaller the mean absolute error, the higher the probability of selection.

The single point crossover method is adopted as the breeding method. The most common breeding method is to generate a child by selecting one crossing point on the parent chromosome and exchanging chromosomal parts before and after the point. For example, if the parent chromosome is given as (24, 262, 343, 789, 1290), (38, 454, 554, 1322, 1449), and there is a crossing point between the third and fourth elements, then the child chromosomes are (24, 262, 343, 1322, 1449), (38, 454, 554, 789, 1290).

When the offspring are created, they have a chance of mutating a portion of their chromosomes, which randomly replaces several elements with completely new values. This new information does not exist in the current population. It is a process that makes it possible to search to compensate for the case where there is no proper solution in the initial set combination.

In this way, a new generation of populations are created by replacing some or all of the existing populations with newly obtained entities. Repeat this process until the number of generations reaches a predetermined value (usually chosen between 10 and 1000), and then selects and ends with the regression model with the best predictive performance.

Once this best multiple linear regression model has been selected, the next step is to examine its validity. If a problem is found that the t-test value of the molecular descriptors included in the model is not good, go back and look for another model. For example, if the number of sample compounds is 1005 and the selected model consists of five molecular descriptors, if the t-test for one of the molecular descriptors is 3.3 or higher, then the probability that the molecular descriptor is irrelevant to that property is 0.1 It means less than%. In the present invention, when there is a molecular presenter having a t-test value of less than about 3, the selected model is discarded and another model is found. In addition, even if the values of the molecular descriptors for the sample compounds are the same except for a few few compounds, they are not considered to be reliable models. In general, increasing the number of molecular expressions included in the model increases the predictive performance, but such problems occur. Therefore, the final model is usually obtained by repeating these steps through several trials and errors while changing the number of molecular expressions. If the problem no longer appears in the selected model, proceed to the next step.

Next, in Step 7, assess the predictive performance of the found model using test sets that did not participate in model formation. If a problem is found in the training set that results in much lower predictive performance or samples that are significantly off predicted, go to step 4 and readjust the training and test sets before proceeding. If the difference between the training set and the test set does not exceed 20% of the absolute mean error (AAE) obtained for the training set, it is judged that the predictive performance is satisfied.

In this way, once a multiple linear regression model for the lower volume percent of flammable is established, the first step is to normalize the data of molecular presenters and experimental data of the lower volume percent of flammability to establish an artificial neural model, that is, the mean of the data at each value. Subtract and divide by the standard deviation. The entire sample thus prepared is divided into a training set, a validation set, and a test set in an approximately 6: 2: 2 ratio.

을, 출력층의 활성화 함수로는 선형함수 즉

를 채택하였다. 따라서 입력층의 각 노드들이 받는 입력값들을

라 할 때 은닉층의 j번째 노드의 출력값은

와 같이 주어지며 은닉층이

개의 노드로 이루어져 있을 때 출력층 출력노드의 최종 출력값은

와 같이 주어진다. 여기서

는 문턱 가중치(threshold weight)를 의미한다.We then use them to find the best artificial neural network model. In this case, the search range is limited to a neural network having a hidden layer between the input layer and the output layer as shown in FIG. 2 and having three structures connected only in a feed forward direction, that is, in a direction from the input to the output. It was. The input layer consisted of the same number of nodes that received the values of each of the molecular markers included in the established polylinear regression model, and the output layer consisted of one node that outputs the lower volume percentage of the printout. In addition, as the activation function of the hidden layer, the Sigmoid function,

, The activation function of the output layer is a linear function

Was adopted. Therefore, the input values that each node in the input layer receives

In this case, the output value of the j th node of the hidden layer is

Is given by

The final output value of the output layer output node when composed of four nodes

As shown in Fig. here

Denotes a threshold weight.

들의 다양한 초기값세트(보통 1000세트이내)를 마련하고, 훈련집합을 사용하여 각 세트로 초기화된 신경망을 역전파 알고리즘을 통해 반복 훈련함으로써 가중치

들의 최적화된 값을 찾는다. 최적화에 대한 판단은 매 훈련 후 경신된 가중치들의 값으로 정해지는 모형을 검증집합에 적용하였을 때 그 평균제곱오차(mean square error)의 값이 최소가 되는 것으로 한다. 보통은 3000~5000번의 반복훈련 내에 이러한 시점이 나오게 된다. 이렇게 얻어진 각 초기값세트에 대응하는 최적화된 신경망모형을 훈련집합, 검증집합, 시험집합에 각각 적용하여 그 평균제곱오차들이 모두 다중선형회귀모형의 그것들보다 작은 것만을 모은다. 이러한 것이 여러 개 있을 경우, 결정계수나 평균절대오차 등을 기준으로 가장 우수한 모형을 선택한다.The search proceeds from increasing the number of hidden nodes to one in order. Usually, the search proceeds to twice the number of input nodes. However, if a satisfactory model is not found, the search proceeds further. The detailed procedure is as follows. First, the weight generated by using random number generation function for each number of hidden nodes

Prepare different sets of initial values (usually within 1000 sets) and weight them by repeatedly training the neural networks initialized with each set using the back-up algorithm.

Find the optimal value for these. The judgment of the optimization is that the mean square error is minimized when the model, which is determined by the updated weights after each training, is applied to the test set. Normally this will occur within 3000 to 5000 repetitions. The optimized neural network model corresponding to each set of initial values thus obtained is applied to the training set, the test set, and the test set, respectively, to collect only those whose mean square errors are smaller than those of the multiple linear regression model. If there are several of these, choose the best model based on the coefficient of determination or the absolute absolute error.

When the artificial neural network model is selected, an overfitting prevention standard is finally set. As a result of over-training, the artificial neural network sets the reference value as a measure to improve the instability of the wrong input to the unknown input. Determine if the prediction will benefit, and choose a reference value (0.4 vol% for hydrocarbons and 1.5 vol% for non-hydrocarbons) that maximizes the gain. When the absolute value of the difference between the predicted values of the artificial neural network model and the multiple linear regression model exceeds the reference value, the predicted value of the multiple linear regression model is adopted.

The results of establishing the QSPR model for the lower flammable volume percentage through this process are summarized in Tables 1 and 2.

Highlights from the QSPR Prediction Model for Lower Flammability Volume Percent of Hydrocarbons Number of sample compounds 356 Number of molecular descriptors 10 Names of Molecular Presenters P1: H autocorrelation of lag 3 / Weighted by atomic Sanderson electronegativities
P2: RPCG Relative positive charge (QMPOS / QTPLUS)
P3: 3D Balaban index
P4: Moment of inertia C
P5: average eigenvector coefficient sum from distance matrix
P6: Relative number of double bonds
P7: Average bond order of a H atom
P8: minimum bond order of carbon atoms (> 0.1) (Min (> 0.1) bond order of a C atom)
P9: Balaban X index
P10: bond information content (neighborhood symmetry of 0-order) Regression Model Decision Coefficients 0.982895 Regression Model AAE 0.041137 vol% Regression model

(n=10)Lower Flammability Volume Percent (vol%) =

( n = 10) Artificial Neural Network Determination Coefficient 0.991641 Artificial Neural Network AAE 0.029814vol% Artificial Neural Network Model Lower Flammability Volume Percent (vol%)
=

( l = 10, m = 20) Overconformity Prevention Criteria 0.4vol%

Key Contents of QSPR Prediction Model for Lower Flammability Volume Percent of Non-Hydrocarbon Number of sample compounds 834 Number of molecular descriptors 15 Names of Molecular Presenters P1: 1 / molecular weight
P2: 3rd component accessibility directional WHIM index / weighted by atomic electrotopological states
P3: Presence of phase distance 2 CC coupling
P4: 3rd component accessibility directional WHIM index / weighted by atomic masses
P5: number of Oxygen atoms
P6: Average valency of a H atom
P7: rotatable bond fraction
P8: Relative number of triple bonds
P9: Gravitation index (all bonds)
P10: R autocorrelation order 1 / unweighted
P11: Relative number of O atoms
P12: Relative number of rings
P13: Min partial charge for a C atom
P14: Narumi's Geometric Topological Index
P15: Average combined information content (degree 0) Regression Model Decision Coefficients 0.945949 Regression Model AAE 0.180697vol% Regression model

(n=15)Lower Flammability Volume Percent (vol%) =

( n = 15) Artificial Neural Network Determination Coefficient 0.973789 Artificial Neural Network AAE 0.121171 vol% Artificial Neural Network Model Lower Flammability Volume Percent (vol%)
=

( l = 15, m = 10) Overconformity Prevention Criteria 1.5vol%

3 is a parity diagram showing the predictive performance of each model, and it can be seen from these figures that the QSPR model has excellent performance. In addition, the average absolute error of the experimental data of the 1190 compounds of which the experimental error is known is about 0.25 vol%, and the error between the experimental value and the predicted value is plotted as a histogram based on this value. In these figures, the QSPR model demonstrates that the QSPR model is accurate by showing that it has a 92.35% probability of predicting the lower percentile volume percentage within the range of the mean experimental error.

The present invention is not limited to the above-described embodiments, and any person having ordinary skill in the art to which the present invention pertains may make various modifications without departing from the gist of the present invention as claimed in the claims. Such changes are intended to fall within the scope of the claims.

Claims (28)

A first step of inputting hydrocarbon-based experimental data among collected sample organic compounds;
A second step of preparing a molecular presenter value for the lower limit of volume of flammability of the hydrocarbon-based organic compound input in the first step;
Extracting optimal molecular descriptors;
A fourth step of separating the experimental data into a training set and a test set;
A fifth step of searching for an optimal multiple linear regression model for the training set;
A sixth step of examining the validity of the selected model;
A seventh step of repeating the fifth and sixth steps if there is no validity in the sixth step, and testing the predictive performance of the model with respect to the test set if it is valid;
An eighth step of repeating steps 4 to 7 if the performance does not satisfy the criterion in the seventh step test for the test set, and separating the three sets after sample normalization if the criterion is satisfied;
A ninth step of searching for an optimal artificial neural network model after dividing the entire sample into three sets;
The absolute value of the difference between the lower flammable volume percent predicted value obtained by the optimal multilinear regression model satisfying the performance test in the eighth step and the lower flammable volume percent predicted value obtained by the optimal neural network model found in the nineth step. A tenth step of comparing with a preset overfit prevention reference value; And
If the difference is greater than the overfit reference value, the lower limit volume percent predicted by the multiple linear regression model obtained in the eighth step is used as the lower limit of the volume percentage, and if the difference is smaller than the overfit reference value, the search is performed in the ninth step. A method for obtaining a lower flammable volume percentage of a hydrocarbon-based organic compound through a multiple linear regression-artificial neural network model comprising the eleventh step of adopting the lower flammable volume percentage predicted value of the artificial neural network model as the lower flammable volume percentage value.
The method of claim 1, wherein the optimal molecular descriptors in the third step are independent molecular descriptors whose values are not the same for all the sample compounds, and ignition of the hydrocarbon-based organic compounds through the multiple linear regression-artificial neural network model. How to find the lower volume percentage.
The method of claim 1, wherein the training set and the test set in the fourth step is divided by the ratio of 5: 5 to 8: 2, the lower limit of the flammable volume percentage of the hydrocarbon-based organic compound through a multiple linear regression-artificial neural network model How to obtain.
2. The multiple linear regression-artificial neural network model according to claim 1, wherein in the fifth step, the multiple linear regression model searches for a multiple linear regression model by applying a genetic algorithm to the training set. Lowering the volume percent lower limit of flammability of hydrocarbon-based organic compounds.
The method of claim 4, wherein the genetic algorithm generates a population composed of a plurality of multiple linear regression models randomly drawn from a predetermined number of molecular representations in a pool of molecular representations. Making; Encoding each individual by combining the numbers of the extracted molecular presenters; Selecting two parent chromosomes from the created population by the Roulette Wheel method and generating offspring by a single point crossover method; And generating a new population by mutating a portion of the chromosomes of the generated progeny with a certain probability, and then replacing a part of the existing population with them to generate a new population. A method for obtaining the lower volume percent of flammable organic compound.
The method of claim 1, wherein the fifth step is to calculate the lower flammable volume percentage of the hydrocarbon-based organic compound through the multiple linear regression-artificial neural network model including determining the predictive performance by the coefficient of regression or the mean absolute error. Way.
The method of claim 1, wherein the lower limit of the flammable volume percentage of the hydrocarbon-based organic compound is obtained through a multilinear regression-artificial neural network model in which the validity in the sixth step is determined by a t-test value.
The method of claim 1, wherein in the eighth step, if the predictive performance of the test set is similar to the predicted performance of the training set, a multiple linear regression model is determined, and the predictive performance of the test set is different from the predicted performance of the training set. The method then calculates the lower percent flammable volume of hydrocarbon-based organic compounds through a multiple linear regression-artificial neural network model that reclassifies the training and test sets.
The method of claim 1, wherein in the ninth step, the search range by the neural network has one hidden layer between the input layer and the output layer and is connected only in a feed forward. A method of obtaining the lower flammable volume percentage of a hydrocarbon-based organic compound through neural network model.
10. The method of claim 9, wherein the ignition lower limit volume percentage of the hydrocarbon-based organic compound is obtained through a multiple linear regression-artificial neural network model, using a sigmoid function as an activation function of the hidden layer.
2. The method of claim 1, wherein the threshold value for flammability of the hydrocarbon-based organic compound is obtained through a multiple linear regression-artificial neural network model according to claim 10, wherein the reference value for preventing overfitting is 0.4 vol%.
A first step of inputting non-hydrocarbon-based experimental data among collected sample organic compounds;
A second step of preparing a molecular presenter value for a lower ignition limit volume percentage of the non-hydrocarbon based organic compounds of the sample organic compounds;
Extracting optimal molecular descriptors;
A fourth step of separating the experimental data into a training set and a test set;
A fifth step of searching for an optimal multiple linear regression model for the training set;
A sixth step of examining the validity of the selected model;
A seventh step of repeating the fifth and sixth steps if there is no validity in the sixth step, and testing the predictive performance of the model with respect to the test set if it is valid;
An eighth step of repeating the fourth to seventh steps if the performance is not satisfied in the seventh step test on the test set, and separating the three sets after normalizing the sample if the performance is satisfied;
A ninth step of searching for an optimal neural network model after dividing the entire sample into three sets;
The absolute value of the difference between the lower flammable volume percent predicted value obtained by the optimal multilinear regression model satisfying the performance test in the eighth step and the lower flammable volume percent predicted value obtained by the optimal neural network model found in the nineth step. A tenth step of comparing with a preset overfit prevention reference value; And
If the difference is greater than the overfit reference value, the lower limit volume percent predicted by the multiple linear regression model obtained in the eighth step is used as the lower limit of the volume percentage, and if the difference is smaller than the overfit reference value, the search is performed in the ninth step. A method for obtaining a lower flammable volume percentage of a non-hydrocarbon-based organic compound using a multiple linear regression-artificial neural network model comprising the eleventh step of adopting a lower flammable volume percentage predicted value by the artificial neural network model as a lower flammable volume percentage value.
13. The method of claim 12, wherein the optimal molecular descriptors in the third step are independent molecular descriptors whose values are not the same for all the sample compounds. How to find the lower flammable volume percentage.
The lower flammable volume percentage of the non-hydrocarbon-based organic compound through the multiple linear regression-artificial neural network model of claim 12, wherein the training set and the test set are divided in a ratio of 5: 5 to 6: 4 in the fourth step. How to obtain.
13. The method of claim 12, wherein in the fifth step, the multiple linear regression model searches for the multiple linear regression model by applying a genetic algorithm to the training set. How to find the lower percent flammable volume of non-hydrocarbon organic compounds.
The method of claim 15, wherein the genetic algorithm generates a population composed of a plurality of multiple linear regression models randomly drawn from the pool of molecular presenters. Making; Encoding each individual by combining the numbers of the extracted molecular presenters; Selecting two parent chromosomes from the created population by the Roulette Wheel method and generating offspring by a single point crossover method; And generating a new population by mutating a portion of the chromosomes of the generated progeny with a certain probability and replacing a part of the existing population with them to generate a new population. A method of obtaining the lower volume percentage of flammable organic compounds.
The method of claim 12, wherein the fifth step is to determine the lower flammable volume percentage of the non-hydrocarbon-based organic compound through the multiple linear regression-artificial neural network model including determining the predictive performance by the regression model crystal coefficient or the mean absolute error. How to obtain.
The method of claim 12, wherein the lower limit of percent flammable volume of the non-hydrocarbon-based organic compound is obtained through a multiple linear regression-artificial neural network model in which the validity in the sixth step is determined by a t-test value.
The method of claim 12, wherein in the eighth step, if the predictive performance of the test set is similar to the predicted performance of the training set, a multiple linear regression model is determined, and the predictive performance of the test set is different from the predicted performance of the training set. After that, the lower limit of percent flammability of the normal state of the non-hydrocarbon organic compounds is obtained through a multiple linear regression-artificial neural network model that reclassifies the training set and the test set.
The method of claim 12, wherein in the ninth step, the search range by the artificial neural network has one hidden layer between the input layer and the output layer and is connected only in a feed forward. How to find the lower percent of flammable volume of non-hydrocarbon organic compounds by neural network model.
21. The method of claim 20, wherein the activation function of the hidden layer uses a sigmoid function to obtain a lower flammable volume percentage of the non-hydrocarbon-based organic compound through a multiple linear regression-artificial neural network model.
13. The method of claim 12, wherein the threshold value of flammability of the non-hydrocarbon-based organic compound is obtained through the multiple linear regression-artificial neural network model, wherein the reference value for preventing overfitting is 1.5 vol%.
The method of claim 1, wherein the optimal molecular descriptors extracted in the third step
P1: H autocorrelation of lag 3 / Weighted by atomic Sanderson electronegativities,
P2: RRCG Relative positive charge (QMPOS / QTPLUS),
P3: 3D Balaban index,
P4: Moment of inertia C,
P5: average eigenvector coefficient sum from distance matrix,
P6: Relative number of double bonds,
P7: Average bond order of a H atom,
P8: minimum bond order of carbon atoms (> 0.1) (Min (> 0.1) bond order of a C atom),
P9: Balaban X index and
P10: A method for obtaining the lower flammable volume percentage of a hydrocarbon-based organic compound using a multiple linear regression-artificial neural network model that includes bond information content (neighborhood symmetry of 0-order).
The method of claim 12, wherein the optimal molecular descriptors extracted in the third step is
P1: 1 / molecular weight,
P2: 3rd component accessibility directional WHIM index / weighted by atomic electrotopological states,
P3: presence / absence of C-C at topological distance 02,
P4: 3rd component accessibility directional WHIM index / weighted by atomic masses,
P5: number of Oxygen atoms,
P6: Average valency of a H atom,
P7: rotatable bond fraction,
P8: Relative number of triple bonds,
P9: Gravitation index (all bonds),
P10: R autocorrelation of lag 1 / Unweighted,
P11: Relative number of O atoms,
P12: Relative number of rings,
P13: Min partial charge for a C atom,
P14: Narumi's Narumi geometric topological index and
P15: Method of obtaining lower flammable volume percentage of non-hydrocarbon-based organic compounds through a multiple linear regression-artificial neural network model containing average bonding information content (order 0).
In the method of obtaining the lower flammable volume percentage of a hydrocarbon-based organic compound through a multiple linear regression-artificial neural network model,
P1: H autocorrelation of lag 3 / Weighted by atomic Sanderson electronegativities,
P2: RRCG Relative positive charge (QMPOS / QTPLUS),
P3: 3D Balaban index,
P4: Moment of inertia C,
P5: average eigenvector coefficient sum from distance matrix,
P6: Relative number of double bonds,
P7: Average bond order of a H atom,
P8: minimum bond order of carbon atoms (> 0.1) (Min (> 0.1) bond order of a C atom),
P9: Balaban X index and
P10: Lower flammability limit volume percentage of hydrocarbon-based organic compounds in a multiple linear regression-artificial neural network model comprising bond information content (neighborhood symmetry of 0-order). How to obtain.
In the method of determining the lower percent of flammable volume of non-hydrocarbon organic compounds by multiple linear regression-artificial neural network model,
P1: 1 / molecular weight,
P2: 3rd component accessibility directional WHIM index / weighted by atomic electrotopological states,
P3: presence / absence of C-C at topological distance 02,
P4: 3rd component accessibility directional WHIM index / weighted by atomic masses,
P5: number of Oxygen atoms,
P6: Average valency of a H atom,
P7: rotatable bond fraction,
P8: Relative number of triple bonds,
P9: Gravitation index (all bonds),
P10: R autocorrelation of lag 1 / Unweighted,
P11: Relative number of O atoms,
P12: Relative number of rings,
P13: Min partial charge for a C atom,
P14: Narumi's Narumi geometric topological index and
P15: A method for obtaining the lower flammable volume percentage of a non-hydrocarbon-based organic compound through a multiple linear regression-artificial neural network model comprising average bonding information content (order 0).
A computer-readable storage medium having recorded thereon a program for executing on a computer a method for obtaining the lower flammable volume percentage of a hydrocarbon-based organic compound according to any one of claims 1 to 11, 23 and 25. .
A computer-readable storage program for executing on a computer a method for obtaining the lower flammable volume percentage of a non-hydrocarbon based organic compound according to any one of claims 12 to 22, 24 and 26. media.

Images