Abstract: Sponge titanium and titanium dioxide are the main products of the titanium metallurgy industry. Titanium slag is the key raw material for the preparation of sponge titanium and titanium dioxide, and its preparation method is the high-temperature reduction smelting of ilmenite in an electric furnace. The high-temperature reduction smelting process of ilmenite has many characteristics different from the ordinary pyrometallurgical smelting process. During this process, the iron oxide in the ilmenite is selectively reduced to metallic iron, and the titanium oxide is enriched in the slag. The by-product of metal iron and the main product of titanium slag are obtained by the separation of molten iron and slag. However, there are some problems in the reduction smelting process of ilmenite, such as low reaction rate, poor separation of slag and iron, and inferior quality of titanium slag. The control of the transport properties of titanium slag melt is key to achieve the efficient preparation of high-quality titanium slag. In this work, the initial configuration for the first-principles molecular dynamics simulation was constructed using classical molecular dynamics. According to the calculation results of the first-principles molecular dynamics simulation, the dataset was constructed, and the accurate machine learning potential function was trained based on the neural network theory. The local structure and transport properties of the TiO₂–FeO–Ti₂O₃ system was studied by machine learning molecular dynamics simulation. The results show that the average bond lengths of Ti⁴⁺—O^2?, Ti³⁺—O^2?, and Fe²⁺—O^2? are 1.88, 1.88, and 1.83 ?, respectively. There are two main connection modes between TiO_n^m? polyhedral units: corner-sharing and edge-sharing. The TiO₆^8? and TiO₆^9? octahedra are involved in the construction of network skeleton. The stability of TiO₆^8? octahedron is higher than that of TiO₆^9? octahedron. Under different FeO contents, TiO₆^8? and TiO₆^8? are the main octahedra in the system. When the mass fraction of FeO increases from 5% to 19%, the variation of the average CN (coordination number) value is Ti³⁺—O^2? > Ti⁴⁺—O^2?; the tricluster oxygen and bridge oxygen in the system are transformed into nonbridge oxygen and free oxygen; the DSC (degree of structure complexity) value of the system decreases from 1.37 to 0.62; Q₄, Q₅, and Q₆ are transformed into Q₀, Q₁, Q₂, and Q₃; and the DOP (degree of polymerization) value decreases from 4.34 to 1.84. The sequence of diffusion abilities of different ions is displayed as follows: Fe²⁺ ≈ O^2? > Ti³⁺ > Ti⁴⁺. When the mass fraction of FeO increases from 5% to 19%, the complexity and polymerization degree of the system and overall strength of the network skeleton decreases, and the viscosity value of the system decreases from 0.043 Pa·s to 0.037 Pa·s. In this work, the correlation model between the viscosity value and the structural parameter DSC value of TiO₂–FeO–Ti₂O₃ system was constructed. The model can reveal the root cause of the change of the viscosity of the system from the physical essence and predict the viscosity of the system. The results will lay the theoretical and technical foundation for the low-carbon and efficient preparation of high-quality titanium slag.