Multi-Head Attention + Soft Actor-Critic for Porosity Prediction

Kianoush Aqabakee and Leonardo Stella submitted a paper to arXiv on 18 June 2026 describing a continuous-action Soft Actor-Critic agent augmented with a multi-head attention feature extractor for porosity prediction and process-parameter optimization.

The paper validates the approach on laser powder bed fusion, reporting that the proposed method achieves a convergence value of 322.79 within 14 episodes and outperforms standard reinforcement learning baselines including DQN, PPO, TD3, and vanilla SAC while maintaining training stability.

What did the authors do?

They replaced a conventional feature extractor with a multi-head attention module inside a Soft Actor-Critic framework and trained the agent in a continuous action space for porosity prediction and process parameter optimization in laser powder bed fusion. The paper contrasts this design with traditional RL approaches that rely on discrete action spaces, which the authors say suffer from slow convergence and susceptibility to local optima.

The experiments report faster convergence and higher final reward values for the attention-enhanced SAC agent. The submission lists DQN, PPO, TD3, and vanilla SAC as comparison methods and states the proposed method reached a convergence value of 322.79 within 14 episodes.

How does the architecture differ from prior RL methods?

The core difference is twofold: moving to a continuous action space and integrating a multi-head attention-based feature extractor with Soft Actor-Critic. According to the paper, the attention extractor improves the agent's ability to capture subtle variations in low-dimensional input features, which helps navigation of value spaces that contain local minima.

The abstract frames discrete action spaces in prior RL work as prone to slow convergence and local optima, positioning the continuous-action SAC plus attention architecture as a way to improve exploration and exploitation balance. The authors state the combined design leads to both faster convergence and greater stability during training compared to the named baselines.

Why it matters

Porosity is a critical defect mode in laser powder bed fusion, and process-parameter optimization directly affects part quality. By reporting a convergence value of 322.79 in 14 episodes and claiming superior final rewards to DQN, PPO, TD3, and vanilla SAC, the paper signals that combining attention mechanisms with continuous-action RL can materially change optimization speed and training stability for this manufacturing task.

Faster, more stable convergence reduces the computational and experimental budget needed to tune process parameters, which matters for researchers and engineers running costly print experiments or simulation suites.

What to watch

Look for the paper's full results and code links referenced on the arXiv entry to validate reproducibility and to see numeric comparisons beyond the single convergence figure provided. Also watch for follow-up work testing the attention-SAC combination across other additive manufacturing metrics or alternative process types.

References: the submission is titled "Multi-Head Attention-Based Feature Extractor Integration with Soft Actor-Critic for Porosity Prediction and Process Parameter Optimization in Additive Manufacturing" by Kianoush Aqabakee and Leonardo Stella, arXiv:2606.20087, submitted 18 Jun 2026.