Abstract: The challenges regarding the lack of integrated spatio-temporal representation of multi-source heterogeneous data, data fragmentation, and the "black box" phenomenon of the melting process in vacuum suspension melting equipment under extreme working conditions are addressed in this paper. Consequently, a digital twin model construction method based on Unreal Engine 5 (UE5) is proposed. Firstly, a process behavior model that accurately maps equipment action sequences is constructed by employing a "feature-motion" joint lightweight strategy and a function-structure decoupling method. Secondly, a physical reduced-order model of the melting process is derived based on the lumped parameter method, and a dynamic mapping mechanism linking physical state parameters with the Niagara particle system is established, through which the visual reconstruction of the melt's microscopic evolution is realized. Validation results indicate that the full process flow is accurately executed by the model, achieving high-fidelity dynamic characterization spanning from macroscopic mechanical motion to microscopic metal phase transitions. Compared with traditional modeling methods, the polygon count of the model is reduced by 34.3%, and the system operating frame rate is increased by 35.1%. The contradiction between high-fidelity mechanism characterization and real-time interactive performance in industrial digital twins is effectively reconciled by this method, providing a visual model foundation for high-level digital twin applications such as intelligent operation and maintenance of complex equipment.