基于迁移学习的大尺寸铸锭晶体硅热场设计

摘要/Abstract

摘要： 不同尺寸的铸锭晶体硅生长过程具有相似性,小尺寸晶体的生长规律可以迁移至大尺寸。本文采用迁移学习(TL)对G8型铸锭炉进行热场设计,设计对象为侧、顶加热器位置及体积、侧隔热笼分区块高度,主要设计目标为减少晶体内部的位错缺陷、抑制硅锭边缘多晶且使晶体生长界面微凸。首先使用神经网络对已有的G7铸锭炉建立热场几何参数与热场评价参数间的映射模型,然后将该模型迁移至G8铸锭炉,对比不同模型结构对迁移过程的影响,采用Dropout分析模型是否存在过拟合,并使用遗传算法(GA)结合聚类算法(CA)对热场几何参数进行优化,以上为G8热场设计过程。最后对优化结果采用数值模拟方法研究其在晶体生长过程中的温度分布、固液界面形状等,最终选定的优化方案能够实现较高质量的长晶。将该方案同时应用于G7和G8热场并进行对比,结果表明G8在硅熔体和硅晶体中的轴向温度梯度均小于G7,在晶体生长界面沿径向的温度梯度也小于G7,这有利于减小晶体内部的热应力。

关键词: G8型铸锭炉, 晶体硅, 热场设计, 迁移学习, 神经网络, 遗传算法

Abstract: There are similarities between the growth processes of ingot crystalline silicon with different sizes, so the growth law of small size crystals could be transferred to large size crystals. In this paper, transfer learning (TL) was used to design the hot zone of the G8 ingot furnace. The design targeted parameters are the positions and volumes of the side and top heaters, and the height of the partition block on the side insulation cage. The main design goals are to reduce the dislocation defects inside the crystal, suppress polycrystalline at the edge of the silicon ingot and make the solid-liquid interface slightly convex. First, a neural network was used to establish a mapping model between the hot zone geometric parameters and hot zone evaluation parameters for the existing G7 ingot furnace. Then the model was transferred to the G8 ingot furnace. The effects of different model structures on the transfer process were analyzed. Then Dropout was used to determine whether the model is overfitting. The genetic algorithm (GA) and the clustering algorithm (CA) were applied to optimize the hot zone geometric parameters. The above is the process of G8 hot zone design. Finally, the numerical simulation method was used to study the temperature distribution and solid-liquid interface shape of the optimized schemes. The final selected scheme could achieve high quality crystal growth. Then the scheme was applied to the hot zones of G7 and G8 furnaces. The results show that the axial temperature gradient of G8 in silicon melt and silicon crystal is smaller than that of G7, and the radial temperature gradient at the solid-liquid interface is also smaller than that of G7. It is beneficial to reduce the thermal stress inside the crystal.

Key words: G8 ingot furnace, crystalline silicon, hot zone design, transfer learning, neural network, genetic algorithm

中图分类号:

郝佩瑶, 郑丽丽, 张辉, 廖继龙. 基于迁移学习的大尺寸铸锭晶体硅热场设计[J]. 人工晶体学报, 2022, 51(8): 1323-1336.

HAO Peiyao, ZHENG Lili, ZHANG Hui, LIAO Jilong. Hot Zone Design of Large Size Ingot Crystalline Silicon Using Transfer Learning[J]. JOURNAL OF SYNTHETIC CRYSTALS, 2022, 51(8): 1323-1336.

参考文献

[1] 王世杰,陈杰勇,江华,等.中国光伏产业发展线路图[R].北京:中国光伏行业协会,2016.
WANG S J, CHEN J Y, JIANG H, et al. Road map of China's photovoltaic industry development[R]. Beijing: China Photovoltaic Industry Association, 2016(in Chinese).
[2] WEI J A, ZHANG H, ZHENG L L, et al. Modeling and improvement of silicon ingot directional solidification for industrial production systems[J]. Solar Energy Materials and Solar Cells, 2009, 93(9): 1531-1539.
[3] MA X, ZHENG L L, ZHANG H, et al. Thermal system design and optimization of an industrial silicon directional solidification system[J]. Journal of Crystal Growth, 2011, 318(1): 288-292.
[4] MA W C, ZHONG G X, SUN L, et al. Influence of an insulation partition on a seeded directional solidification process for quasi-single crystalline silicon ingot for high-efficiency solar cells[J]. Solar Energy Materials and Solar Cells, 2012, 100: 231-238.
[5] YU Q H, LIU L J, MA W C, et al. Local design of the hot-zone in an industrial seeded directional solidification furnace for quasi-single crystalline silicon ingots[J]. Journal of Crystal Growth, 2012, 358: 5-11.
[6] 娄中士,左然,苏文佳,等.大晶粒多晶硅铸锭生长的热场设计与模拟[J].人工晶体学报,2011,40(6):1602-1606.
LOU Z S, ZUO R, SU W J, et al. Thermal system design and simulation for growth of large grain multi-crystal silicon ingot[J]. Journal of Synthetic Crystals, 2011, 40(6): 1602-1606(in Chinese).
[7] 陆晓东,张鹏,吴元庆,等.定向凝固多晶硅铸锭炉石英坩埚的改进与热场优化[J].人工晶体学报,2015,44(11):3179-3183.
LU X D, ZHANG P, WU Y Q, et al. Thermal field optimization and improvement of directional solidification of polycrystalline silicon ingot furnace quartz crucible[J]. Journal of Synthetic Crystals, 2015, 44(11): 3179-3183(in Chinese).
[8] 滕冉,戴小林,徐文婷,等.热屏优化对大直径单晶硅生长影响的数值模拟[J].人工晶体学报,2012,41(1):238-242+252.
TENG R, DAI X L, XU W T, et al. Numerical simulation on effect of heat-shield optimization on the growth of large-diameter silicon single crystal[J]. Journal of Synthetic Crystals, 2012, 41(1): 238-242+252(in Chinese).
[9] 张向宇,关小军,潘忠奔,等.热屏位置对直拉硅单晶V/G、点缺陷和热应力影响的模拟[J].人工晶体学报,2014,43(4):771-777.
ZHANG X Y, GUAN X J, PAN Z B, et al. Simulation on effect of heat shield position on the V/G and point defect and thermal stress of Czochralski silicon[J]. Journal of Synthetic Crystals, 2014, 43(4): 771-777(in Chinese).
[10] YU W C, ZHU C, TSUNOOKA Y, et al. Geometrical design of a crystal growth system guided by a machine learning algorithm[J]. CrystEngComm, 2021, 23(14): 2695-2702.
[11] PAN S J, YANG Q. A survey on transfer learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345-1359.
[12] YAMADA H, LIU C, WU S, et al. Predicting materials properties with little data using shotgun transfer learning[J]. ACS Central Science, 2019, 5(10): 1717-1730.
[13] JHA D, CHOUDHARY K, TAVAZZA F, et al. Enhancing materials property prediction by leveraging computational and experimental data using deep transfer learning[J]. Nature Communications, 2019, 10: 5316.
[14] KUDO H, MATSUMOTO T, KUTSUKAKE K, et al. Occurrence prediction of dislocation regions in photoluminescence image of multicrystalline silicon wafers using transfer learning of convolutional neural network[J]. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2021, E104.A(6): 857-865.
[15] 王晋东,陈益强.迁移学习导论[M].北京:电子工业出版社,2021.
WANG J D, CHEN Y Q. Introduction to transfer learning[M]. Beijing: Publishing House of Electronics Industry, 2021(in Chinese).
[16] 郝佩瑶,朱金伟,廖继龙,等.基于神经网络和遗传算法的铸锭晶体硅质量控制及工艺优化[J].人工晶体学报,2022,51(3):385-397.
HAO P Y, ZHU J W, LIAO J L, et al. Process control and optimization of ingot crystalline silicon growth using neural network and genetic algorithm[J]. Journal of Synthetic Crystals, 2022, 51(3): 385-397(in Chinese).
[17] 马旭.晶硅定向凝固过程中的热输运与组分输运研究[D].北京:清华大学,2012.
MA X. Thermal and species transport during directional solidification of silicon for solar cells[D]. Beijing: Tsinghua University, 2012(in Chinese).
[18] SRIVASTAVA N, HINTON G, KRIZHEVKY A, et al. Dropout: A simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014, 15: 1929-1958.
[19] 玄光男,林林.网络模型与多目标遗传算法[M].梁承姬,于歆杰,译.北京:清华大学出版社,2017:151.
GEN M, HAYASHI L. Network model and multi-objective genetic algorithm[M]. LIANG C J, YU X J, Transl. Beijing: Tsinghua University Press, 2017: 151(in Chinese).
[20] MACQUEEN J. Some methods for classification and analysis of multivariate observations[C]//Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. 1967, 1(14): 281-297.
[21] CALIŃSKI T, HARABASZ J. A dendrite method for cluster analysis[J]. Communications in Statistics, 1974, 3(1): 1-27.
[22] ZIO E, BAZZO R. A clustering procedure for reducing the number of representative solutions in the Pareto Front of multiobjective optimization problems[J]. European Journal of Operational Research, 2011, 210(3): 624-634.