Attention-Based SAC Optimizes Additive Manufacturing Parameters

Optimizing additive manufacturing processes is crucial for minimizing defects like porosity, which requires precise control over process parameters. Traditional reinforcement learning (RL) methods, often relying on discrete action spaces, struggle with slow convergence and susceptibility to local optima, limiting their effectiveness in high-precision manufacturing. This research addresses these limitations by proposing a novel architecture that combines a continuous action space with a multi-head attention mechanism integrated into the Soft Actor-Critic (SAC) algorithm. The attention-based feature extractor significantly improves the agent's ability to discern subtle variations within low-dimensional input features. This enhanced feature understanding allows for a more effective balance between exploration and exploitation, which is vital for navigating complex value spaces that contain multiple local minima. The proposed methodology was validated on tasks involving porosity prediction and process parameter optimization in laser powder bed fusion. The results demonstrate faster convergence and higher final reward values compared to several standard RL algorithms, including DQN, PPO, TD3, and a vanilla SAC implementation. The system achieved a convergence value of 322.79 within just 14 episodes, showcasing its superior performance and stability throughout the training process.