Multimodal perception fusion control for Czochralski crown growth in a full-scale furnace

Abstract

The existing controllers are stuck in simulation, lack of applicability for nonlinearity uncertainty, and time variation of Czochralski crown growth. To address the problem, a multimodal perception fusion control (MPFC) method was developed to predict the pulling speed, thereby effectively controlling diameter of the crown growth. To enhance the generalizability, MPFC integrates cross-modal knowledge mining and crystal growth kinetics to ensure the quality of crown growth and minimize the invalid output of the controller. MPFC achieved high accuracy on the test dataset, with an R-squared of 0.87 and a root mean square error (RMSE) of 3.64. MPFC was compared with ResNet + temporal convolutional network, proportional-integral-derivative controllers, and physics-informed neural networks using RMSE. Furthermore, its high accuracy, adaptability, and generalizability were validated through control simulations and experiments in a full-scale furnace under a wide range of initial states. MPFC is a robust method for industrial crown growth, enabling precise diameter control and maximized survival rates.