Variation of the slag cover surface (SCS) in the oxygen-enriched top-blown molten bath smelting process is critical for the smelting efficiency of a complex Cu–S concentrate. However, capturing these variation characteristics is difficult because of the high temperature inside the molten bath and the dynamic complexity of the smelting process. In this work, machine learning (i.e., U-net algorithm and support vector machine) is combined with a skillful hydraulic model (i.e., gas–liquid two-phase top-blown agitated vessel) and an experimental measurement strategy to quantitatively explore the variation characteristics of the SCS in an oxygen-enriched top-blown molten bath smelting process. Results showed that a minimum of 30 images, with the smallest size being 900 × 600 pixels, was sufficient for the training process. The data accuracy of the training procedure ranged from 93.20% to 96.23% for identifying the SCS at the laboratory scale. The highest average height of 2.23 cm for the SCS occurred under the operational condition, with a flow rate of 160 L/h, a liquid temperature of 60 °C, and a liquid depth of 0.4 m. The chaotic systems of SCS in industry were deterministic. It was found that the proposed strategy could be used to accurately identify the variation characteristics of the SCS in the gas–liquid two-phase top-blown agitated vessel. The variation of the SCS in the industrial process could be roughly grasped by magnifying the height of the SCS obtained from the experimental data in the laboratory. Quantification of the variation characteristics of the SCS is useful to increase the smelting efficiency of the oxygen-enriched top-blown molten bath smelting process. This also provides insights for multiphase measurements in other studies related to efficient utilization of complex Cu–S concentrates.