Abstract: "Integral forming and mirror milling" represents an inevitable trend in the manufacturing of high-performance propellant tank domes. Issues such as unstable cutting vibration, out-of-control wall thickness accuracy, invalid positioning datum, and mismatched compensation response have long been critical bottlenecks restricting the mirror milling of a certain type of closed integral ultra-thin tank dome with a diameter of 3350 mm and a wall thickness of 1 mm. To address these problems, an overcut prediction method for mirror milling of ultra-thin tank domes based on long short-term memory (LSTM) neural networks is proposed. By analyzing the time series of parameters strongly correlated with overcut in ultra-thin tank domes, a dual-channel overcut prediction model is established using LSTM neural networks and rule-based judgment. Engineering verification shows that partitioned machining and partitioned data acquisition, combined with the construction of model training datasets and dynamic overcut evaluation indicators, enable the prediction of overcut-related parameters at the next time step and real-time overcut risk assessment during ultra-thin tank dome machining. This method has supported the successful completion of mirror milling for China's first ultra-thin tank dome with a diameter of 3350 mm and wall thickness of 1 mm, with the overall wall thickness meeting the accuracy requirement of ±0.1 mm.