简介概要

Boron removal from metallurgical grade silicon by slag refining based on GA-BP neural network

来源期刊：Rare Metals2021年第1期

论文作者：Shi-Lai Yuan Hui-Min Lu Pan-Pan Wang Chen-Guang Tian Zhi-Jiang Gao

文章页码：237 - 242

摘要：In order to investigate the boron removal effect in slag refining process,intermediate frequency furnace was used to purify boron in SiO₂-CaO-Na₃ AlF₆-CaSiO₃ slag system at 1,550℃,and back propagation（BP） neural network was used to model the relationship between slag compositions and boron content in SiO₂-CaO-Na₃ AlF₆-CaSiO₃ slag system.The BP neural network predicted error is below 2.38 %.The prediction results show that the slag composition has a significant influence on boron removal.Increasing the basicity of slag by adding CaO or Na₃ AlF₆ to CaSiO₃-based slag could contribute to the boron removal,and the addition of Na₃ AlF₆ has a better removal effect in comparison with the addition of CaO.The oxidizing characteristic of CaSiO₃ results in the ineffective removal with the addition of SiO₂.The increase of oxygen potential（pO₂）in the CaO-Na₃ AlF₆-CaSiO₃ slag system by varying the SiO₂ proportion can also contribute to the boron removal in silicon ingot.The best slag composition to remove boron was predicted by BP neural network using genetic algorithm（GA）.The predicted results show that the mass fraction of boron in silicon reduces from 14.0000 × 10^-6 to0.4366×10^-6 after slag melting using 23.12 % SiO₂-10.44 % CaO-16.83 % Na₃ AlF₆-49.61 % CaSiO₃ slag system,close to the experimental boron content in silicon which is below 0.5×10^-6.

稀有金属(英文版) 2021,40(01),237-242

Boron removal from metallurgical grade silicon by slag refining based on GA-BP neural networkP neural network predicted error is diction results show that the slag ficant influence on boron removal.f slag by adding CaO or Na3AlF6 to d contribute to the boron removal,AlF6 has a better removal effect in The development of solar energy,increase of environmental conscidevelopment,increases the deman(SoG-Si).In the past decades,deenvironmentally friendly method

Shi-Lai Yuan Hui-Min Lu Pan-Pan Wang Chen-Guang Tian Zhi-Jiang Gao

School of Materials Science and Engineering,Beijing University of Aeronautics and Astronautics

作者简介：Shi-Lai Yuan e-mail:yshi10303@163.com;Hui-Min Lu e-mail:13331151800@126.com;

收稿日期：22 October 2013

基金：financially supported by the National High Technology Research and Development Program of China (No.2012AA062302);

Boron removal from metallurgical grade silicon by slag refining based on GA-BP neural network

Shi-Lai Yuan Hui-Min Lu Pan-Pan Wang Chen-Guang Tian Zhi-Jiang Gao

School of Materials Science and Engineering,Beijing University of Aeronautics and Astronautics

Abstract：

In order to investigate the boron removal effect in slag refining process,intermediate frequency furnace was used to purify boron in SiO₂-CaO-Na₃ AlF₆-CaSiO₃ slag system at 1,550℃,and back propagation(BP) neural network was used to model the relationship between slag compositions and boron content in SiO₂-CaO-Na₃ AlF₆-CaSiO₃ slag system.The BP neural network predicted error is below 2.38 %.The prediction results show that the slag composition has a significant influence on boron removal.Increasing the basicity of slag by adding CaO or Na₃ AlF₆ to CaSiO₃-based slag could contribute to the boron removal,and the addition of Na₃ AlF₆ has a better removal effect in comparison with the addition of CaO.The oxidizing characteristic of CaSiO₃ results in the ineffective removal with the addition of SiO₂.The increase of oxygen potential(pO₂)in the CaO-Na₃ AlF₆-CaSiO₃ slag system by varying the SiO₂ proportion can also contribute to the boron removal in silicon ingot.The best slag composition to remove boron was predicted by BP neural network using genetic algorithm(GA).The predicted results show that the mass fraction of boron in silicon reduces from 14.0000 × 10^-6 to0.4366×10^-6 after slag melting using 23.12 % SiO₂-10.44 % CaO-16.83 % Na₃ AlF₆-49.61 % CaSiO₃ slag system,close to the experimental boron content in silicon which is below 0.5×10^-6.

Keyword：

Metallurgical grade silicon; Boron removal; Slag system; Genetic algorithm-back propagation neural network;

Received： 22 October 2013

1 Introduction

The development of solar energy,which is affected by the increase of environmental consciousness and the societal development,increases the demand on solar grade silicon(SoG-Si).In the past decades,developing a low cost and environmentally friendly method to produce SoG-Si triggered intensive research.Refining of metallurgical grade silicon (MG-Si) was found to be an alternative method in comparison with the traditional Siemens process.As such,impurities in MG-Si must be removed to low levels for the photovoltaics (PV) cell to operate at optimum efficiency.Impurities with its higher vapor pressure than silicon can be evaporated preferentially by vacuum refining process.Zheng et al. [ 1] found that phosphorus content in MG-Si could decrease to less than 0.1 x 10^-6 after vacuum treatment.However,this method is ineffective in boron removal,for the saturated vapor pressure of boron is 6.78×10^-7 Pa,which is far less than that of silicon (0.4 Pa).Directional solidification can segregate impurities from solid to the melt,which means that the first solidification part is pure.The removal efficiency depends on the difference between impurity concentrations of solid and melt,which is defined as segregation coefficient.Unfortunately,the segregation coefficient of boron is 0.8,close to 1,resulting in the invalid removal of boron using this method [ 2, 3, 4, 5] .

Slag refining process is an effective method to separate boron from silicon.The mechanism of slag refining process is considered to include two main steps:the oxidation of boron and the absorption of boron oxide by slag.The oxidation of boron,described as the oxygen potential (pO2),can be achieved by silica resulting from the equilibrium Si and SiO_2.The absorption of boron oxide,described as the slag basicity,is expected to facilitate their extraction to alkali and alkali-earth oxides in the slag phase.Thus,the removal of boron is expected to be strongly dependent on the slag chemistry.Leandro et al. [ 6] examined the effect of boron removal in CaO-SiO_2-based ternary systems and showed that the addition of excess CaO can decrease the activity of SiO_2.Johnston and Barati [ 7] ,Wu et al. [ 8] ,Cai et al. [ 9] ,Luo et al. [ 10] ,and Ding et al. [ 11] studied the effects of CaO-SiO_2-Al₂O₃,CaO-SiO_2-Na₂O,Al₂O_3-BaO-SiO₂,CaO-SiO_2-Li₂O,and Al₂O₃-CaO-MgO-SiO₂ slag systems on boron removal.Li₂O,Na₂0,MgO,and BaO are thought to be associated with B₂O_3.Alumina behaves as a network-forming oxide by joining the silicate chain structure within slag.Owing to the depolymerization of silicate network by F^-which increases the concentration of free O2^-,the addition of CaF₂ will increase the basicity of the CaO-SiO₂ slag system.

Artificial neural network (ANN) is an algorithm mathematical model dealing with distributed parallel information and characterized by imitating the structure and functional aspects of biological neural network.Owing to their excellent ability of non-linear mapping,self-organization and self-learning,ANN is widespread in engineering area.The BP neural network is a kind of typical‘feed-forward,back propagation'neural network,trained by back-propagation of errors (BPNN),according to the training of the multilayer network [ 12] .Genetic algorithm (GA) is a calculation model by simulating the Darwin's evolution process through natural selection and genetic mutation to obtain the better weights and bias.BP algorithm refrains the network from trapping in a local minimum [ 13] .

The objective of the present work is to study the boron removal effects in slag refining process using SiO_2-CaO-Na₃ AlF_6-CaSiO₃ slag system and to present a new method of applying the technology of neural network to forecast the boron removal model in order to make a comprehensive understanding of the factors that affect the boron removal in the slag refining process.Detailed GA-BP neural network training and prediction process was given.Furthermore,the behavior of slag composition in boron removal was analyzed.The slag compositions were arranged using the uniform design method in order to have a larger coverage.Then,the relationship between slag compositions and boron removal was built using BP neural network model.

2 Experimental

SiO₂,CaO,Na₃AlF₆,and CaSiO₃ powders,used in this experiment,were classified as fine powders.The MG-Si ingot with the boron content of 14×10^-6 was prepared.The slag compositions were designed using the uniform design method,with the proportion of SiO₂,CaO,Na₃AlF₆varying from 0%to 30%,respectively.The detailed slag compositions are shown in Table 1.The powders used in the preparation of the slag were dried at 100℃for 1 h in order to remove humidity completely,then weighed and mixed thoroughly in the desired ratio for each particular experiment.The slag powder mixtures were pressed into a pellet.50 g silicon ingot and 50 g slag mixture were used in each test.The slag pellet and slag ingot were added to an alumina crucible which was inside a graphite crucible in an intermediate frequency furnace.The atmosphere in the furnace was below 0.01 Pa before the raw mixtures were heated up.Then,the temperature of the mixture was heated up to 1,550℃which was measured and also controlled by an infrared thermometer and kept for 1 h.A schematic diagram of the intermediate frequency furnace is shown in Fig.1.

After slag refining for 1 h,the electric power was turned off,and the molten phase was cooled down along the water-cooled copper coil.The slag and the metal phases were physically separated from the crucible.The test samples were cut from the center of each ingot,then ground to powder.The boron content in silicon was analyzed by inductively coupled plasma atomic emission spectroscopy (ICP-AES) and inductively coupled plasma mass spectrometry (ICP-MS),respectively.

下载原图

Table 1 Compositions of slags before equilibrium experiments(wt%)

Fig.1 Schematic diagram of intermediate frequency furnace

3 Results and discussion

3.1 GA-BP neural network training and prediction process

An ANN consists of an interconnected group of artificial neurons,and it processes information using a connectionist approach to computation.In our prediction process,there are four inputs (SiO₂ proportion,CaO proportion,Na₃AlF₆proportion,and CaSiO₃ proportion) and one output (boron content),and the ANN with three layers was chosen to construct the non-linear reflection model,as shown in Fig.2.The common ANN is a feed-forward network.The weights and the threshold values of ANN are randomly generated.During the ANN training process,back-propa?gation of errors was used to decrease the training error and improve the training efficiency.The training epochs and training goal are 2,000 and 1×10^-8,respectively.

To increase the predicted accuracy and training efficiency,the GA was used to optimize BP neural network structure,transfer function,weights,and the threshold values by selecting best fitness chromosome,and crossing with their father generation and mutation.The flowchart of GA-BP neural network is shown in Fig.3.The selected operation of GA process firstly uses the elitist strategy,copies the top 10%fitness inpidual of father generation to the new generated population,and uses roulette-wheel algorithm to select the rest of chromosome.This method can avoid local optimal gene which dominates the whole generation,while global optimum would likely be eliminated before emerging.Crossover operation is a process that exchanges the independently distributed random variables with the range of [ 1] in two different chromosomes.Mutation operation changes the values of one gene in the selected chromosome,and this process can make GA obtain local random research capability.The parameters of the GA were set as following:the initial population number N=100,the cross probability P_c=0.8,the mutation probability P_m=0.08,and the error e=1×10^-8.

Fig.2 Common artificial three-layer neural network

Fig.3 Flowchart of GA-BP neural network

In order to improve training accuracy,the sample data received from slag refining experiments need normalization pretreatment.Then input values and network training output values both in binary coding have the same regularity,which makes the convergence faster in network training and the prediction relatively accurate.The GA-BP neural network program was processed under Matlab2013a.The predicted output and error percentages are shown in Fig.4.The largest predicted error percentage is 2.38%.The low error percentage explains that the application of the GA-BP neural network model in forecasting the boron removal effect is feasible and reliable.The GA-BP neural network prediction model shows that the boron content in silicon reduces from 14.0000 x 10^-6to 0.4366×10^-6 after slag melting in 23.12%SiO_2-10.44%CaO-16.83%Na₃AlF_6-49.61%CaSi0₃ slag system.The prediction results are verified by the experimental data,which shows that the boron content could decrease to<0.5×10^-6 under the same experimental condition when the prediction slag composition was used.This can also verify the validity of the GA-BP neural network.

Fig.4 Operation results of GA-BP neural network mode:a forecast output and actual output,and b prediction error percentage

3.2 Ternary slag on boron removal

The effects of CaSiO_3-based ternary slag composition on boron removal were calculated as shown in Fig.5.It is found that the boron content in silicon ingot decreases with the increase of Na₃AlF₆ proportion in slag and reaches a minimum value of 1.28×10^-6 when the Na₃AlF₆ proportion is 0.17.Then the boron content in silicon ingot increases when the Na₃AlF₆ proportion exceeds 0.17.It is concluded that the depolymerization of silicate network by F^-increases the concentration of free O^2-.CaO in slag plays a positive role in boron removal,which is expected to combine with (BO₃)^3-and form CaO·B₂O₃ compound.The boron removal effect in Na₃AlF_6-CaSiO₃ slag system is better than that in CaO-CaSiO₃ slag system.This is because the viscosity and the melting temperature of the former slag is lower than that of the latter slag,for the melt temperature of Na₃AlF₆ (1,000℃) is below that of CaO(2,615℃) [ 14] .It will increase the activity coefficient of slag components to make the slag have a better effect on boron removal.Furthermore,the (BO₃)^3-combines with Na₂O instead of CaO on account of the more negative standard Gibbs free energy of the formation of Na₂O·B₂O₃than that of CaO·B₂O₃ as shown in Fig.6,with calculated data from thermochemical data of pure substances.The boiling temperature and the decomposition temperature of NaBO₂ are relatively low (1,474.5℃).Boron may escape from the slag-silicon system in the form of NaBO₂ or its decomposition products with the assistance of the volatilization of sodium compounds when the temperature of the silicon-slag system is high enough.The boron content in silicon ingot slightly decreases as the SiO₂ proportion in SiO_2-CaSiO₃ slag increases from 0%to 30%corresponding to the increase of .SiO₂ exists in the slag as an oxidized specie under the conditions employed in the present works.But increasing oxygen potential in slag haslittle effect on boron removal,which indicates that theoxygen potential is not the limit factor on boron removal inthe SiO_2-CaSiO₃ slag system.

Fig.5 Effects of Na₃AlF₆,SiO₂,CaO proportion on boron removal

Fig.6 Standard Gibbs free energies for formation of compounds CaO·B₂O₃ and Na₂O·B₂O₃

3.3 Basicity analysis

The effect of slag basicity on boron removal from silicon was predicted by varying the addition of CaO and Na₃ AlF₆ with CaSiO₃ fixed at 60%.The boron removal effects are shown in Fig.7.The boron content in silicon ingot follows a similar tendency as that in the ternary slags of CaO-CaSiO₃ and Na₃AlF_6-CaSiO₃ in Fig.5.But the boron content is lower than that in the latter slag refining systems because of the CaSiO₃ fixed at 60%and the addition of SiO₂ to slag.This is because that the addition of SiO₂ to the slag can increase the limits of the oxygen partial pressure (pO2) resulting from the equilibrium between Si and SiO₂ [ 15] ,which would decrease as the CaO/Na₃ AlF₆ increases in the slag.Compared with the boron content in silicon ingot using SiO_2-CaO-CaSiO₃ slag,boron content in SiO_2-Na₃AlF_6-CaSiO₃ slag system decreases rapidly.The boron content in silicon ingot achieves the same level in both SiO_2-CaO-CaSiO₃ slag and Si0_2-Na₃AlF_6-CaSiO₃ systems as the composition of CaO/Na₃AlF₆ excesses 40%.The boron removal effects of CaO/Na₃AlF₆ in silicon ingot are also predicted as shown in Fig.8.The predicted results show that the boron removal effects slightly change with the composition of the slag changing.The reason that the removal effects maintain the same level in different slag compositions,as shown in Fig.7,is that the oxygen potential is the main factor for the boron removal.The reaction between Na₃AlF₆ and SiO₂ also attributes to the rapid decrease of boron content using SiO_2-Na₃AlF_6-CaSiO₃ slag,for it can decrease the oxygen potential and increase the basicity of slag [ 16] .

3.4 Oxygen potential analysis

The effect of oxygen potential (po2) of the slag was calculated by varying the SiO₂ proportion with the proportion of Na₃AlF₆ and CaSiO₃ fixed at 15%and 45%,respectively.The removal effects are shown in Fig.9.The boron content in silicon ingot decreases with the increase of SiO₂proportion in slag and reaches a minimum value of0.62×10^-6 when the SiO₂ proportion is around 20%.However,the boron content in silicon ingot increases when the Na₃AlF₆ proportion exceeds 20%.For the decreasing part in Fig.9,the ability of the slag oxidizing boron into B₂O₃ increases with the SiO₂ proportion increasing,and there are enough O^2-to combine with B₂O₃ to form(BO₃)^3-and then enter into slag phase.For the uplifting part in Fig.9,the decreasing CaO proportion results in the low content of 0^2-in slag phase,which hinders the B₂O₃in silicon molten from entering into the slag phase and results in the increasing boron concentration in silicon ingot.

Fig.7 Effects of Na₃AlF₆ and CaO on boron removal in SiO2-Na₃AlF₆/CaO-CaSiO₃ salgs

Fig.8 Boron removal effects in Ca0-Na₃AlF₆-CaSiO₃ slag (propor-tion of CaSiO₃ being fixed at 60%)

Fig.9 Effects of SiO₂ on boron removal in SiO_2-CaO-Na₃AlF_6-CaSiO₃ slag system.Proportions of Na₃AlF₆ and CaSiO₃ being fixed at 15%and 45%,respectively

4 Conclusion

The GA-BP neural network was chosen to build model to study relationships between slag compositions and boron content in SiO_2-CaO-Na₃AlF_6-CaSiO₃ slag system at1,550℃.The largest predicted error is 2.38%,indicating that the model is feasible and reliable.The GA-BP neural network prediction results show that the mass fraction of boron in silicon could reduce from 14.0000×10^-6 to0.4366×10^-6 after slag melting in 23.12%SiO_2-10.44%CaO-16.83%Na₃AlF₆-49.61%CaSi0₃ slag system.The experimental data show that the boron content can decrease to<0.5×10^-6 using the prediction slag compositions,which verifies the feasibility and reliability of the model.

The boron content in silicon ingot is strongly affected by the slag composition.The addition of Na₃AlF₆/CaO to CaSiO₃ slag could decrease the boron content remarkably,while the addition of SiO₂ has slight influence on boron removal,for the basicity of the slag is the main factor.The Na₃AlF_6-CaSi0₃ slag behaves a better boron removal effect than the CaO-CaSiO₃ slag.The addition of CaO of basicity in CaSiO_3-based slag could improve the boron removal effect.Compared with CaO,the addition of Na₃AlF₆ has a better affinity to (BO₃)^3-.The increase of oxygen potential (po₂) of the CaO-Na₃AlF₆-CaSiO₃ slag by varying the SiO₂ proportion can also contribute to the boron removal in silicon ingot.