Multi-Head Attention-Based Feature Extractor Integration with Soft Actor-Critic for Porosity Prediction and Process Parameter Optimization in Additive Manufacturing

Additive manufacturing (AM) processes, specifically laser powder bed fusion (L-PBF), require precise control over process parameters to minimize defects like porosity. Traditional reinforcement learning (RL) approaches often struggle with slow convergence or getting trapped in local optima when navigating the complex, high-dimensional parameter spaces of these manufacturing systems. This paper investigates whether integrating a multi-head attention mechanism into the Soft Actor-Critic (SAC) algorithm can improve the agent's ability to extract meaningful features and optimize these parameters more efficiently.

The researchers developed a novel architecture that replaces standard input processing with a multi-head attention-based feature extractor. This mechanism allows the agent to capture subtle variations in input features by computing self-attention weights, which helps the model understand long-range dependencies within the state space. The agent operates in a continuous action space, allowing for finer control over parameters like laser power, scan speed, and hatch spacing. The framework is trained using a reward function derived from the Eagar-Tsai thermal model, which balances the achievement of desired microstructural properties against energy consumption penalties.

The proposed attention-based SAC model demonstrated superior performance compared to standard RL benchmarks, including DQN, PPO, TD3, and vanilla SAC. By mapping the input domain into more informative features, the agent successfully navigated the value space's local minima, achieving a convergence value of 322.79 within only 14 episodes. The results indicate that the attention mechanism provides a more effective balance between exploration and exploitation, leading to both faster training convergence and higher final reward stability.