<p>High-performance thick-walled tubes are widely applied in marine engineering, with multi-pass flow forming as an efficient manufacturing process for them. However, steep strain gradients and severe material accumulation trigger defects such as fish scaling and annular spallation, and inter-pass deformation heredity complicates defect control. In this study, a process-tailored CNN-LSTM model is established to guide process optimization. First, defect formation mechanisms are examined through experiments and simulations. On this basis, quantitative defect-propensity metrics are defined, and a strain-geometry state matrix is formulated to describe the deformation state of tubes. Second, a CNN-LSTM architecture is developed to model the inter-pass evolution of deformation state. In this model, strain-gradient features are extracted by the CNN, and deformation heredity is captured by the LSTM. Finally, a full-factorial design, driven by the model’s predictions, is employed to quantify the effects of process parameters on defect formation, followed by defect-control experiments to validate the model’s reliability. The results suggest that the CNN-LSTM model accurately predicts defect metric evolution across passes with R² ≥ 0.9697. To sum up, lowering the thinning ratio, elevating the feed rate, and decreasing the forming angle are beneficial for suppressing fish scaling and annular spallation defects.</p>