Abstract: Vacuum levitation melting furnaces offer advantages such as contactless heating, high purity, and low pollution, making them widely used in the preparation of high-performance metal materials. However, the complex nonlinear relationships between multi-physics coupling and process parameters within them make precise control difficult for traditional methods that rely on experience and experimentation. This paper constructs a mechanistic model based on thermodynamics and electromagnetics. By combining historical operating data to modify boundary conditions and parameters, it created a digital twin of the melting process, achieving dynamic modeling and updating that was consistent between the virtual and the real. Furthermore, a physicalinformed machine learning approach was introduced to train an AI-powered process parameter prediction model under physical boundary constraints. The AI-predicted parameters were applied to actual melting, and the resulting data was used to inversely optimize the mechanistic and AI models, forming a closed-loop co-evolutionary frame work. This approach significantly improves the accuracy and prediction reliability of the mechanistic model, enabling adaptive process control and intelligent production. It transcends the limitations of traditional empirical control and purely data-driven models to construct a novel intelligent control paradigm integrating interpretable mechanisms and AI adaptability. This approach provides an effective solution for the digital twinning and intelligent upgrading of high-end vacuum metallurgical equipment, as well as for overall technological advancement in the industry.

Key words: vacuum levitation melting furnace, digital twin, artificial intelligence, process parameter prediction