Research On Data-driven Modeling And Control Of Semiconductor Silicon Single Crystal Growth Process

With the rapid development of advanced technologies such as artificial intelligence,big data,and 5G communication,the demand for cloud computing and terminal electronic products has increased significantly,and the contradiction between market supply and demand has been transmitted from the chip manufacturing field to the upstream semiconductor material manufacturing field.At present,semiconductor materials are still dominated by silicon materials.The strong demand for large-size,high-quality integrated circuit-grade semiconductor silicon single crystal materials puts forward higher requirements for crystal growth process and control.However,semiconductor silicon single crystal growth is a process of mutual support and close combination of equipment and process,with a high technological threshold,multi-subject intersection,strong industry monopoly and other outstanding features.Therefore,this thesis focuses on the key scientific issues in the semiconductor silicon single crystal growth process,and carries out process modeling and process control research to improve the quality of semiconductor silicon single crystal to meet the ever-increasing chip process,which is undoubtedly of great theoretical significance and practical value.The main research work of this thesis are as follows:(1)Research on data-driven prediction modeling of key crystal quality parameters:Aiming at the problem that the key quality parameters such as crystal diameter and melt temperature in the silicon single crystal growth process are difficult to be accurately detected online,a process variable prediction model based on discrete time nonlinear autoregressive with exogenous inputs(NARX)model is established.Firstly,an optimal estimation method of timedelay based on cross-correlation function is proposed,which accurately estimates the timedelay between different process variables,and identifies the order of NARX model based on Lipschitz quotient criterion and model goodness-of-fit.Secondly,a deep learning model based on stacked auto-encoder(SAE)and long short-term memory network(LSTM)is designed.The multivariate prediction of crystal diameter and melt temperature is realized by using the data extraction function of SAE and the dynamic time series prediction ability of LSTM.Then,according to the wavelet packet data decomposition technique,SAE,and LSTM network,a decomposition-integration strategy-based ensemble learning modeling method for key quality parameters is proposed to further improve the prediction accuracy of crystal diameter and melt temperature.Finally,the experimental results show that the proposed two data-driven prediction models have good prediction performance.(2)Research on hybrid modeling of silicon single crystal growth process by fusing data and mechanism:Aiming at the problems of inaccurate mechanism model of silicon single crystal growth process and weak interpretability of data-driven model,two hybrid models based on series/parallel structure are constructed.Firstly,the established single-input-single-output hybrid model is mainly composed of an energy transfer model based on LSTM-HammersteinWiener and a crystal pulling kinetic model with crystal diameter estimation compensation,which realizes the robust prediction of crystal growth rate and the accurate estimation of crystal diameter.Secondly,in the process of developing a multi-input-multi-output hybrid model,an energy transfer model JITL-SAE-ELM(Extreme learning machine,ELM)based on a Just-intime learning(JITL)fine-tuning strategy is proposed to solve the problem of online adaptive updating of the model and to achieve accurate multivariate prediction of melt temperature and crystal growth rate.In addition,the designed multi-input-single-output crystal diameter estimation error compensation model not only compensates the modeling error of the energy transfer model and the unmodeled dynamics of the crystal pulling state space model,but also improves the accurate estimation of the crystal diameter parameters by the hybrid model.Finally,the experimental results show that the two hybrid models can achieve the robust prediction of multiple crystal quality parameters,which can effectively reduce the difficulty of pure mechanism modeling and enhance the role of data-driven modeling in the hybrid model.(3)Research on data-driven predictive control of silicon single crystal growth process:Aiming at the crystal diameter and melt temperature control problems such as serious nonlinearity,large hysteresis,time-varying and uncertainty disturbances in the silicon single crystal growth process,an adaptive nonlinear predictive control method for crystal diameter based on ensemble learning modeling is firstly proposed,and the stability of the controller with constraints is analyzed to achieve accurate control of the crystal diameter.Secondly,the hierarchical control method of constant pulling rate silicon single crystal growth process is proposed.A disturbance rejection model-free adaptive controller(MFAC)based on discretetime extended state observer(ESO)is designed for the inner melt temperature control process to ensure the stability of melt temperature control.Meanwhile,a disturbance rejection modelfree adaptive predictive controller(MFAPC)based on ESO is designed for the outer crystal diameter control process,which realizes the optimization of temperature trajectory and improves the control accuracy of crystal diameter parameters.Then,for the case that there is no realization condition of constant pulling rate crystal growth control,a dual control method of silicon single crystal growth process controlled by pulling rate is proposed.The single variable MFAPC controller of melt temperature based on adaptive ESO and the multivariable extended MFAPC controller of crystal diameter are designed,and their stability is analyzed.Through the coordination mechanism of the two controllers,not only the stable and accurate control of melt temperature and crystal diameter is realized,but also the limitation problem of manual setting of crystal pulling rate trajectory and temperature trajectory is solved.Finally,the experimental results of silicon single crystal growth process control show that the proposed three control methods have good stable and accurate control performance of crystal quality parameters.(4)Research on data-driven silicon single crystal batch process control:Aiming at the problems of unstable batch control of key crystal quality parameters during repeated operation of single crystal furnaces and susceptibility to non-repetitive factors between batches,a disturbance rejection model-free adaptive iterative learning control(MFAILC)method based on iterative extended state observer(IESO)is proposed,and the convergence of batch tracking error is analyzed.By designing a MFAILC controller with disturbance suppression for the crystal diameter and melt temperature control process,the stability and robustness of the crystal diameter and melt temperature control between batches are realized.Then,in order to improve the batch-to-batch control accuracy and output stability of the MFAILC controller based on IESO,a data-driven iterative learning model predictive control(ILMPC)method based on the integral form is proposed,and the bounded convergence of the batch tracking error of the controller is proved,so as to realize the stable and accurate control of the melt temperature and crystal diameter.Finally,the batch control experiments of silicon single crystal growth process show that the proposed two data-driven batch control methods ensure that the control system has learning ability and can realize robust batch control of key crystal quality parameters.