Optimization of process parameters to control diameter fluctuation during the monocrystalline silicon growth by hybrid neural network model and genetic algorithm

Abstract

This study aims to construct a neural network-genetic algorithm collaborative optimization framework through a data-driven approach to break the challenge of diameter fluctuation control under multi-parameter coupling. Based on real-time data from the isothermal stage of industrial single crystal furnaces, the Grey Wolf Optimization algorithm is used to dynamically optimize the weights and hyperparameters of the Multi-Layer Perceptron. Consequently, a high-precision surrogate model of the relationship between process parameters (heater power, crystal pulling speed, crucible lift speed) and diameter fluctuation is established. Meanwhile, the thermal field stability is optimized by the Genetic Algorithm (i.e., a global search of the optimal parameter combination). Verified by factory measurements, the model prediction error is less than 5 %, and the optimal solutions for heater power, crystal pulling and crucible lift speed are 53.64 kW, 1.308 mm/min, and 0.163 mm/min, respectively. Actual production shows that the average diameter fluctuation of the crystal rods in the optimization group decreased from 1.48 ± 0.23 mm (original process) to 0.60 ± 0.15 mm (mean ± standard deviation), a reduction of 59.4 %. The coupled method significantly enhances the intelligence level of the single crystal silicon growth process, and provides a generalizable optimization paradigm for the preparation of other semiconductor crystals.