Method for predicting viscosity of metallurgical slag based on microstructure and machine learning

Abstract

The invention discloses a method for realizing metallurgical slag viscosity prediction based on a microstructure and machine learning, and belongs to the technical field of metallurgical slag viscosity prediction. The method comprises the steps that according to specific components of a slag system, Lammps software is adopted for molecular dynamics simulation, and the types and the number of microstructure units of slag are obtained; secondly, establishing a microstructure unit model of the slag by adopting machine learning to obtain the molar content of a structure unit; and finally, inputting the molar content of the structural unit of the to-be-predicted slag obtained from the microstructure model into the deduced and modified slag viscosity prediction equation to obtain a predicted value of the viscosity of the to-be-measured slag. According to the method, the interference of external environmental factors is avoided through a microstructure and machine learning method, the viscosity of the metallurgical slag under different components and different temperature conditions can be quickly and accurately obtained, the precision and stability of predicting the viscosity of the metallurgical slag are remarkably improved, and the cost is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of metallurgical slag viscosity prediction, and in particular to a method for realizing intelligent prediction of metallurgical slag viscosity based on microstructure. Background Art

[0002] Steelmaking is like slag making. Slag plays an important metallurgical function in the modern steelmaking process. From blast furnace and converter refining to continuous casting, slag has an important impact on metallurgical quality, smelting efficiency and smooth metallurgical operation. With the development of steelmaking towards high quality, high efficiency, green and safe direction, higher requirements are put forward for the metallurgical function of metallurgical slag. Viscosity, as an important physical property of slag, directly affects the performance of its metallurgical function. Accurate viscosity helps to understand the dynamic conditions of slag. Viscosity is not only related to the smooth progress of the smelting process, but also affects the heat transfer and mass transfer in the process. Slag with too high viscosity will not only lead to poor fluidity, reduced slag foaming ability, and limit the contact area between slag and molten steel, but also reduce the reaction rate, affect the efficiency of desulfurization and dephosphorization, and make it difficult to fully separate impurities. On the contrary, slag with too low viscosity may not be able to effectively wrap impurities, easily cause splashing, and seriously erode the furnace lining, reducing the life of the furnace lining. Therefore, understanding and controlling slag viscosity is of great significance for optimizing steel production processes and improving product quality.

[0003] At present, there are two main methods to obtain viscosity data: experimental research and model calculation. In the process of experimental research, the viscosity of metallurgical slag is mostly tested by high-temperature viscometer, such as the rotating cylinder method. Although experimental viscosity research can provide test results that are closer to the real environment, its high cost, high energy consumption, high requirements, operational complexity and potential safety risks are all factors that need to be considered. Viscosity prediction models can be divided into theoretical models, empirical models, and semi-theoretical and semi-empirical models based on the construction principle. Model calculation can effectively improve R&D efficiency, reduce costs, and provide valuable reference for actual production. Common viscosity models include the classic Arrhenius model, VFT model, Adam-Gibbs model, Iida model, Mills model, etc. These calculation models are optimized based on limited experimental data, and their scope of application is greatly limited. At the same time, the model parameters are only fitted by a large amount of experimental data, lacking practical physical meaning.

[0004] The macroscopic properties of the melt are determined by the microstructure, and the viscosity can be reflected from a microscopic perspective as the interaction between the internal structural groups of the melt. Metallurgical slag belongs to silicate melt. The oxides in the slag can be divided into network formers, network modifiers and network intermediates according to their contribution to the polymerization degree of silicate melt. The oxygen in the slag can be divided into bridging oxygen (BO), non-bridging oxygen (NBO) and free oxygen (FO). The ratio of bridging oxygen to network formers, NBO / T, can characterize the degree of depolymerization of the slag structure. The value of NBO / T is proportional to the degree of depolymerization of the slag and inversely proportional to the viscosity of the slag. The slag is a network structure formed by [SiO4]-tetrahedrons and [AlO4]-tetrahedrons connected by bridging oxygen, and Al 3+ With Si 4+ When the charge number is different, additional charge compensation tetrahedron is required to maintain the stability of the structure. Alkaline oxides CaO and MgO will decompose into O 2- and the corresponding cations, which act as network modifiers to compensate for the charge, dissociating O 2- It can cut off the bridging oxygen bond of the local structure to form non-bridging oxygen or free oxygen. Non-bridging oxygen represents O 2- Only one endpoint is connected to Si 4+ connection, while free oxygen represents O 2- Neither of the two endpoints is aligned with Si 4+ The bond formed by the endpoint connected to the non-network-forming cation is weak and easy to break, which is regarded as the cut-off point. The oxygen bonds adjacent to the cut-off point will move under the action of shear force, destroying the network structure. The movement of the cut-off point will cause the viscous flow of the melt.

[0005] Molecular dynamics is an effective computational tool for obtaining the structural properties of slag from a microscopic perspective. Through simulation calculations, the short-range and medium-range structures of microscopic particles in slag under different compositions, temperatures and pressures can be directly analyzed, thereby obtaining the types and quantities of microstructures in high-temperature metallurgical slag. The advantages of strong flexibility and economic efficiency make molecular dynamics simulation widely used in the field of metallurgy. However, there are still some shortcomings in molecular dynamics. First, the time scale of particle movement in molecular dynamics simulation is short, usually 1 femtosecond. It may take several days to achieve a meaningful simulation time, which makes molecular dynamics simulation have time scale limitations. Secondly, molecular dynamics simulation requires a lot of computing resources, and the amount of data generated is usually large, containing a lot of complex information, which limits the application of molecular dynamics simulation in large systems or long time scales. With the advancement of computer hardware and intelligent algorithms, artificial intelligence has developed rapidly. As a branch of artificial intelligence, machine learning has been continuously expanding its application scenarios in the industrial field. This technology can better meet the control accuracy requirements. It can effectively and accurately determine various process parameters, which can well promote the automation and intelligent development of the steel industry. Machine learning technology can analyze data information, extract useful information, fit the complex relationship between data variables, and achieve fast and accurate prediction of target data. Therefore, combining molecular dynamics simulation with machine learning to obtain slag microstructure information under continuous composition can make up for the shortcomings of molecular dynamics simulation.

[0006] Therefore, it is of great significance to establish a viscosity prediction model with high prediction accuracy based on microstructure and machine learning and the nature of slag viscosity, which has important application value for the stable operation of industrial production. Summary of the invention

[0007] (I) Purpose of the invention

[0008] In view of the problems existing in the above-mentioned technology, the purpose of the present invention is to provide a method for predicting the viscosity of metallurgical slag by utilizing microstructure and machine learning, which can quickly and accurately obtain the viscosity of metallurgical slag under different compositions and different temperature conditions, and significantly improve the accuracy and stability of metallurgical slag viscosity prediction.

[0009] (II) Technical solution

[0010] In order to achieve the above object, the present invention provides a method for predicting the viscosity of metallurgical slag using molecular dynamics simulation, comprising the following steps:

[0011] S1. First, according to the specific composition of the slag system, the Lammps software is used to perform molecular dynamics simulation, and then the type of structural unit and its molar content of the slag are obtained using Matlab software;

[0012] S2. Using machine learning methods to establish a prediction model for the molar content of structural units of the slag system;

[0013] S3. Inputting the molar content of the slag structural unit to be predicted obtained from the prediction model into the pre-trained slag viscosity prediction model to obtain a predicted value of the slag viscosity to be measured.

[0014] Furthermore, in step S1, the components of the metallurgical slag to be tested include SiO2, Al2O3, CaO and MgO.

[0015] Furthermore, in the step S1, the in file of the Lammps software running command executed also includes boundary conditions, equilibrium ensemble, temperature control method, pressure control method, cooling rate, summation method and relaxation time.

[0016] Furthermore, the boundary conditions adopt periodic boundary conditions, the ensemble is an NVT canonical ensemble, the temperature control method is a Nose-Hoover heat bath method, the pressure control method is a Parrinello-Rahman method, and the summation method is an Ewald summation method.

[0017] Furthermore, in step S1, the equilibrium structure data is read using statistical analysis software Matlab, the atomic information corresponding to each time step is recorded, and each type of atom is distinguished, and the distance between each oxygen atom, silicon atom and aluminum atom in the slag is calculated.

[0018] Furthermore, with the oxygen atom as the center and the cutoff distance of the Si-O bond and the Al-O bond as the radius, the silicon atoms and aluminum atoms around a single oxygen atom are counted. Then, the number of bridging oxygen atoms around each silicon atom (aluminum atom) is determined to obtain the structural unit Q Si n and Q Al n The molar content of .

[0019] Furthermore, in step S2, the CatBoost algorithm based on symmetric decision tree is selected to establish the microstructure model of metallurgical slag, which uses a binary decision tree as a basic predictor. The decision tree is a model established by recursively dividing the feature space into multiple non-overlapping regions (tree nodes) according to the values ​​of certain split attributes, which is used to identify that certain features exceed a certain threshold. Each final region (leaf of the tree) is assigned a value, which is the predicted value of y in the regression task area. The decision tree h can be written as:

[0020]

[0021] Where J is the number of decision tree leaves, b i is the weight of decision tree leaf i, is the indicator function of the decision tree leaf, indicating that when the input x meets all the conditions of the decision tree, the prediction result is the value of leaf i. Further, in step S2, the Pearson correlation coefficient is used for feature selection:

[0022]

[0023] In the formula, r is the characteristic correlation coefficient, which indicates the linear correlation between the characteristic variable and the residual manganese content at the end point, n is the number of samples, and x is i is the true value of each feature, is the average value of each feature, y i is the true value of the molar content of the structural unit, It is the average value of the molar content of the structural unit.

[0024] Further, in the step S2, in the Python language environment and the Jupyter lab development environment, the Visual Studio Code platform is used to perform data processing based on the Scikit-learn, lightgbm and catboost frameworks using Numpy, Pandas and other data processing libraries. The data is standardized using the StandardScaler of the preprocessing preprocessing module in the sklearn framework. The model training set and the test set are divided using the train_test_split function of the preprocessing preprocessing block. The training set is then used for model training, and the test set is used to evaluate the accuracy of the trained model, thereby obtaining a trained microstructure model.

[0025] Furthermore, in step S3, the pre-trained slag viscosity prediction model includes:

[0026] η=A·exp(E η / RT) (1.17)

[0027] lnA=a+bE η (1.18)

[0028]

[0029] Among them, A is the pre-exponential factor, E η is the activation energy of metallurgical slag (J / mol), E v is the activation energy of pure SiO2 and E v =5.21×105 J / mol, R is the gas constant and R=8.314 J / (mol·K), T is the temperature K, a and b are model parameters, α is the parameter of bond strength and α iThe value is between 1.8 and 4, i is an oxide other than SiO2, j represents a non-network-forming cation, represents the mole fraction of j-compensated [AlO4]-tetrahedron, N O 2- is the molar content of free oxygen in the slag, which is assumed to be 0 in this model.

[0030] Furthermore, the model parameter a is based on the binary slag system composed of oxides and SiO2 in the slag. i Obtained,

[0031] a=∑ i (x i a i ) / ∑ i x i (1.20)

[0032] The model parameter b is based on the binary slag system b composed of oxides and SiO2 in the slag. i Obtained,

[0033] b=∑ i (x i b i ) / ∑ i x i (1.21)

[0034] Among them, x i Indicates the molar content of oxide i other than SiO2.

[0035] The non-bridging oxygen and free oxygen produced by CaO, MgO and Al2O3 in the slag can be obtained by subtracting the bridging oxygen number from the total oxygen number, and the molar numbers of the non-bridging oxygen and bridging oxygen produced by each of them can be distributed accordingly according to the proportion of each component in the slag.

[0036]

[0037] (III) Beneficial effects

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention performs molecular dynamics simulation on the melt structure of slag with different contents, determines the type and molar content of microstructure units, and then uses machine learning to establish a microstructure model to obtain the molar content of the microstructure units of the slag. Based on the Arrhenius equation, according to the flow mechanism of the silicate network structure, combined with the type of microstructure units in the melt structure, a viscosity prediction equation that depends on the change of oxygen bonds is established, and finally the nonlinear equation solving function of MATLAB software is used to solve the viscosity prediction equation. Therefore, the present invention proposes for the first time a method for predicting the viscosity of metallurgical slag based on microstructure and machine learning, and realizes the prediction of slag viscosity under a wider range of test conditions through the microstructure of the slag under finite molecular dynamics simulation, which can provide a reference for designing slag with reasonable smelting performance.

[0040] The present invention proposes a viscosity prediction method for metallurgical slag. According to the Arrhenius equation, based on microstructure and machine learning, a viscosity prediction equation that depends on the change of oxygen bonds is established. The constructed viscosity prediction model can efficiently calculate the viscosity value under specific conditions, eliminating tedious and complicated experimental steps, and the method can predict the viscosity of slag under different composition conditions with high accuracy, and has great practicality and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention is described with the aid of the following drawings:

[0042] Figure 1 This is a flow chart of a method for predicting the viscosity of metallurgical slag based on microstructure and machine learning according to Example 1 of the present invention;

[0043] Figure 2 The figure is a relationship curve diagram between the predicted viscosity value and the measured viscosity value according to Example 1 of the present invention. DETAILED DESCRIPTION

[0044] In order to better understand the technical solution of the present invention, the content of the present invention includes but is not limited to the specific implementation methods described below, and similar technologies and methods should be considered to be within the scope of protection of the present invention. In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0045] It should be clear that the embodiments described in the present invention are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "an" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0047] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0048] Example 1

[0049] The present invention selected 20 groups of metallurgical slags published in foreign high-level SCI journals, and the specific components are shown in Table 1:

[0050] Table 1 Composition of experimental slag of CaO-SiO2-Al2O3-MgO system (mass fraction, %)

[0051]

[0052]

[0053] The molar ratio is obtained based on the mass ratio of each component of the slag (CaO, SiO2, Al2O3 and MgO), and then the number of different atoms (Si, Al, Ca, Mg and O) is determined based on the total number of atoms of 8000. Then the density of the slag is obtained based on experimental tests and the side length of the box is calculated, as shown in Table 2.

[0054] Table 2 Atomic number, density and box side length of experimental slag of CaO-SiO2-Al2O3-MgO system

[0055]

[0056]

[0057] (1) Based on the above atomic types, numbers and slag density, a randomly distributed amorphous structure model is constructed. The initial structure data file and potential function of the slag to be tested are imported into the in file that executes the Lammps software running command. The initial structure of the slag to be tested is energy minimized in the Lammps software, and periodic boundary conditions are used to form a periodic image. The shape of the simulated cubic box remains unchanged, and the NVT ensemble is selected to maintain its stability. The Parrinello-Rahman and Nose-Hoover heat bath methods are used for pressure and temperature control. In terms of the algorithm, the cutoff radius of the short-range force is set to The cutoff diameter is smaller than the side length of any system box. The Ewald algorithm is used for long-range Coulomb force. The running time step is set to 1fs, and the data is saved every 10 steps through the frog leaping algorithm, for a total of 800,000 steps. In terms of temperature setting, the initial temperature is first set to 5000K, and 200,000 steps are run to fully mix the system to eliminate the initial distribution state of the atoms. Then, the temperature is reduced to 1773K through 200,000 steps, and finally, the temperature is continued to relax at 1773K for 400,000 steps, of which the first 20,000 steps are equilibrium relaxation. All data collection is completed within the last 200,000 steps, and finally the coordinate position file of 8,000 atoms in the system is output.

[0058] (2) Matlab software was used to read the equilibrium structure data, record the atomic information corresponding to each time step, distinguish the atoms of different types, calculate the distance between each oxygen atom, silicon atom and aluminum atom, and thus obtain the different structural units Q i The concentration distribution of (where i represents the number of bridging oxygen in the structural unit) is shown in Table 3.

[0059] Table 3 Structural unit Q of CaO-SiO2-Al2O3-MgO system experimental slag i content

[0060]

[0061]

[0062] (3) The CatBoost algorithm based on symmetric decision tree was used to establish the microstructure model of metallurgical slag. In the Python language environment and Jupyter lab development environment, the Visual Studio Code platform was used to process the data based on the Scikit-learn, lightgbm and catboost frameworks using Numpy, Pandas and other data processing libraries. The StandardScaler of the preprocessing module in the sklearn framework was used to standardize the data. The train_test_split function of the preprocessing block was used to divide the model training set and test set. Then the training set was used for model training, and the test set was used to evaluate the accuracy of the trained model, so as to obtain the microstructure prediction model under different components of the trained slag system.

[0063] (4) The molar fraction of the structural unit of the slag to be tested is obtained from the established microstructure prediction model, and is input into the calculation formula of oxygen bonds shown in equations (1.27) to (1.31) to calculate the molar number of non-bridging oxygen and bridging oxygen generated by each component:

[0064]

[0065] The number of moles of each oxygen bond is input into the pre-trained slag viscosity prediction model shown in equations (1.32) to (1.34):

[0066] η=A·exp(E η / RT)(1.32)

[0067] (5) The viscosity values ​​of 58 groups of the above 20 groups of samples at three different temperatures of 1723K, 1773K and 1823K were calculated using the model of the present invention. Figure 2 The comparison between the viscosity values ​​calculated by the model and the experimental values ​​is shown, and the errors are within 15%. The Urbain model and the Nakamoto model widely used in the metallurgical industry are also used for comparison. The results show that the prediction of the model of the present invention is 2.8% higher than that of the Urbain model, and 2.1% higher than that of the Nakamoto model, which verifies the accuracy of the model of the present invention.

[0068] Based on the above analysis, the viscosity prediction method of the present invention is applicable to all metallurgical slags with SiO2, Al2O3, CaO and MgO as main components. The present invention fully considers the influence mechanism of the network structure of silicates on fluidity, combines the type of oxygen bonds in the melt structure, and establishes a viscosity prediction model that depends on the change of oxygen bonds.

[0069] The above description shows and describes the preferred embodiments of the present invention, but as mentioned above, it should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the application concept described herein through the above teachings or the technology or knowledge of the relevant field. And the changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for predicting the viscosity of metallurgical slag based on microstructure and machine learning, characterized in that: The steps include: S1. First, according to the specific composition of the slag system, the Lammps software is used to perform molecular dynamics simulation, and then the type of structural unit and its molar content of the slag are obtained using Matlab software; S2. Using machine learning methods to establish a prediction model for the molar content of structural units of the slag system; S3. Inputting the molar content of the slag structural unit to be predicted obtained from the prediction model into the pre-trained slag viscosity prediction model to obtain a predicted value of the slag viscosity to be measured.

2. The method according to claim 1, characterized in that In the step S1, the components of the metallurgical slag to be tested include SiO2, Al2O3, CaO and MgO.

3. The method according to claim 1, characterized in that In the step S1, the in file of the Lammps software running command also includes boundary conditions, equilibrium ensemble, temperature control method, pressure control method, cooling rate, summation method and relaxation time.

4. The method according to claim 3, characterized in that The boundary conditions adopt periodic boundary conditions, the ensemble is the NVT canonical ensemble, the temperature control method is the Nose-Hoover heat bath method, the pressure control method is the Parrinello-Rahman method, and the summation method is the Ewald summation method.

5. The method according to claim 1, characterized in that: In step S1, the equilibrium structure data is read using the statistical analysis software Matlab, the atomic information corresponding to each time step is recorded, and each type of atom is distinguished to determine the type of the formed structural unit.

6. The method according to claim 5, characterized in that With the oxygen atom as the center and the cutoff distance of the Si-O bond and Al-O bond as the radius, the silicon atoms and aluminum atoms around a single oxygen atom are counted. Then the structural unit Q is determined by determining the number of bridging oxygens around each silicon atom (aluminum atom). n Type.

7. The method according to claim 1, characterized in that In step S2, the Boosting integration algorithm in machine learning is used to allow the secondary learner to fit the prediction error of the previous learner, and then iterative training is performed. Finally, the prediction results of these learners are weighted and combined to generate an accurate structural unit prediction model, thereby obtaining a database of the molar content of each structural unit within the slag system.

8. The method according to claim 1, characterized in that In step S2, the Boosting type integrated algorithm is a supervised classification learning method, and the CatBoost algorithm is selected to establish the microstructure model of metallurgical slag. CatBoost is an implementation of gradient boosting, which uses a binary decision tree as a basic predictor. The decision tree h can be written as:

9. The method according to claim 1, characterized in that: In step S2, the training of the machine learning model requires feature selection, that is, selecting the optimal subset of variables that have an impact on or have predictive power on the target variable from all variables in the data set. The Pearson correlation coefficient is defined as the quotient of the product of the covariance and standard deviation between two variables, which is used to measure the degree of linear correlation between two variables: In the formula, r is the characteristic correlation coefficient, which indicates the linear correlation between the characteristic variable and the residual manganese content at the end point, n is the number of samples, and x is i is the true value of each feature, is the average value of each feature, y i is the true value of the molar content of the structural unit, It is the average value of the molar content of the structural unit.

10. The method according to claim 1, characterized in that In step S3, the pre-trained slag viscosity prediction model includes: η=A·exp(E η / RT) (1.4) lnA=a+bE η (1.5) Among them, A is the pre-exponential factor, E η is the activation energy of metallurgical slag (J / mol), E v is the activation energy of pure SiO2 and E v =5.21×105 J / mol, R is the gas constant and R=8.314 J / (mol·K), T is the temperature K, a and b are model parameters, α is the parameter of bond strength and α i The value is between 1.8 and 4, i is an oxide other than SiO2, j represents a non-network-forming cation, represents the mole fraction of j-compensated [AlO4]-tetrahedron, N O 2- is the molar content of free oxygen in the slag, which is assumed to be 0 in this model.

11. The method according to claim 10, characterized in that The model parameter a is based on the binary slag system composed of oxides and SiO2 in the slag. i The model parameter b is obtained based on the b of the binary slag system composed of oxides and SiO2 in the slag. i Obtained, Among them, x i Indicates the molar content of oxide i other than SiO2. The non-bridging oxygen and free oxygen produced by CaO, MgO and Al2O3 in the slag can be obtained by subtracting the bridging oxygen number from the total oxygen number, and the molar numbers of the non-bridging oxygen and bridging oxygen produced by each of them can be distributed accordingly according to the proportion of each component in the slag.

Images