Investigation on Welding Sequence Optimization for Ship Structural Members Using Reinforcement Learning

Welding sequence strongly affects assembly distortion of thin-plate ship structures, but sequence planning still relies on skilled workers' experience and trial-and-error. This study proposes an AI welding-sequence optimization system integrating reinforcement learning with finite-element method (FEM) distortion analysis. The problem is formulated as a sequential decision process where states encode execution history and actions select the next weld line to minimize final deformation. For a four-weld-line specimen, Q-learning is applied and the derived sequences are validated by thermo-elasto-plastic FEM analysis and stereo-vision measurements of out-of-plane displacement and end-angle, confirming good agreement between analysis and experiments. For a ship-block-scale model with 22 weld lines, a Deep Q-Network (DQN) combined with inherent-strain-based FEM enables efficient search and reduces peak out-of-plane displacement from 30~mm to 15~mm. This study demonstrates that reinforcement learning can autonomously extract physically interpretable welding sequence strategies that control history-dependent thermal–mechanical interactions in multi-pass joining processes. The framework provides practical decision support for welding sequence planning in shipbuilding within realistic computation time.