pith. sign in

arxiv: 2606.20087 · v1 · pith:3PRW3YZHnew · submitted 2026-06-18 · 💻 cs.AI

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

Kianoush Aqabakee , Leonardo Stella This is my paper

Pith reviewed 2026-06-26 17:07 UTC · model grok-4.3

classification 💻 cs.AI
keywords additive manufacturingreinforcement learningsoft actor-criticmulti-head attentionporosity predictionprocess parameter optimizationlaser powder bed fusion
0
0 comments X

The pith

Multi-head attention integrated with Soft Actor-Critic reaches a convergence value of 322.79 in 14 episodes for additive manufacturing parameter optimization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes integrating a multi-head attention feature extractor with the Soft Actor-Critic algorithm to optimize continuous action spaces for porosity prediction and process parameters in laser powder bed fusion. Traditional RL methods using discrete actions converge slowly and get stuck in local optima, but the attention component is intended to better capture subtle variations in low-dimensional features. This leads to faster convergence and higher rewards than DQN, PPO, TD3, and vanilla SAC while preserving training stability. A sympathetic reader would care because improved parameter control could reduce defects in high-precision manufacturing.

Core claim

The proposed methodology integrates a multi-head attention-based feature extractor with the Soft Actor-Critic algorithm, achieving a convergence value of 322.79 within 14 episodes on porosity prediction and process parameter optimization in laser powder bed fusion while outperforming DQN, PPO, TD3, and vanilla SAC and maintaining stability throughout training.

What carries the argument

Multi-head attention mechanism serving as a feature extractor inside the Soft Actor-Critic framework to improve capture of subtle variations in low-dimensional inputs and balance exploration-exploitation in continuous action spaces.

If this is right

  • The method converges faster than standard RL approaches in continuous action spaces for manufacturing optimization.
  • It reaches higher final reward values than DQN, PPO, TD3, and vanilla SAC.
  • Training remains stable across episodes while navigating value spaces with local minima.
  • Continuous action spaces become practical for high-precision tasks such as porosity minimization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same attention-augmented SAC structure could be tested on other parameter-optimization tasks that involve low-dimensional sensor data.
  • If the performance gain holds, hybrid attention-RL agents might reduce the number of physical trials needed during process development.
  • Extending the approach to multi-objective rewards that include build time or energy use would be a direct next step.

Load-bearing premise

The multi-head attention mechanism enhances the agent's ability to capture subtle variations in low-dimensional input features, enabling a more effective exploration-exploitation balance.

What would settle it

A controlled ablation that removes the multi-head attention component and measures whether convergence slows below 14 episodes or the final reward falls below 322.79 on the same laser powder bed fusion task would settle the claim.

Figures

Figures reproduced from arXiv: 2606.20087 by Kianoush Aqabakee, Leonardo Stella.

Figure 1
Figure 1. Figure 1: Characterization in production, with different partition [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Value Space related to reward function 1 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-Head Attention Architecture D. Multi-Head Attention-based Soft Actor-Critic To address the optimization problem based on RL, we define our reward function as (1). The policy π drives the selection of actions, and the corresponding action value function Qπ , representing the cumulative reward for a given action under π, is estimated using neural networks. These networks are referred to as the actor an… view at source ↗
Figure 4
Figure 4. Figure 4: Multi-Head Attention-based Soft Actor-Critic Main Diagram [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Return diagram rolling window size 30 proposed algorithm demonstrates the best performance across all criteria. The final aspect to be analyzed is whether the agent successfully converged to the global solution. To provide further insight into the training process, Fig.7 illustrates vari￾ous components of the agent. The return value stabilizes at the convergence point after approximately six episodes, as d… view at source ↗
Figure 5
Figure 5. Figure 5: Return diagram rolling window size 5 presented in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Proposed Multi-Head Attention-based SAC agent re [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
read the original abstract

Additive manufacturing process optimization requires precise parameter control to minimize defects such as porosity. Traditional reinforcement learning (RL) approaches using discrete action spaces suffer from slow convergence and susceptibility to local optima, limiting their effectiveness for high-precision manufacturing tasks. This study addresses these limitations by employing a continuous action space combined with a novel architecture that integrates a multi-head attention mechanism with the Soft Actor-Critic (SAC) algorithm. The attention-based feature extractor enhances the agent's ability to capture subtle variations in low-dimensional input features, enabling more effective exploration-exploitation balance for navigating value spaces with local minima. We validate our approach on porosity prediction and process parameter optimization in laser powder bed fusion, demonstrating faster convergence and higher final reward values compared to standard RL methods including DQN, PPO, TD3, and vanilla SAC. The proposed methodology achieves a convergence value of 322.79 within 14 episodes, outperforming existing approaches while maintaining stability throughout training.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes integrating a multi-head attention mechanism as a feature extractor with the Soft Actor-Critic (SAC) algorithm to optimize process parameters and predict porosity in laser powder bed fusion additive manufacturing. It claims that this yields faster convergence to a value of 322.79 within 14 episodes while maintaining stability, outperforming DQN, PPO, TD3, and vanilla SAC on the task.

Significance. If the numerical claims can be reproduced with fully specified state/reward definitions and controls, the work could demonstrate a practical benefit of attention-augmented continuous RL for manufacturing optimization. The current text supplies no such specification, so the result cannot yet be evaluated for significance or generality.

major comments (3)
  1. [Abstract] Abstract: the reward function (e.g., whether it is negative porosity, a shaped potential, or a weighted sum) and the exact state vector are never defined, rendering the headline scalar result of 322.79 uninterpretable and preventing any meaningful comparison to the listed baselines.
  2. [Abstract] Abstract: no information is supplied on dataset size, number of input features, train/test split, or statistical significance testing, so the claim of outperforming DQN/PPO/TD3/SAC cannot be assessed.
  3. [Abstract] Abstract: the assertion that multi-head attention improves exploration-exploitation balance on low-dimensional features is presented without ablation studies, attention-weight visualizations, or comparison to a non-attention SAC variant, leaving the architectural contribution unsupported.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for greater methodological detail and supporting analyses. We agree these elements are necessary for reproducibility and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reward function (e.g., whether it is negative porosity, a shaped potential, or a weighted sum) and the exact state vector are never defined, rendering the headline scalar result of 322.79 uninterpretable and preventing any meaningful comparison to the listed baselines.

    Authors: We agree that explicit definitions are required for interpretability. In the revised manuscript we will add a dedicated subsection specifying the state vector (laser power, scan speed, hatch spacing, layer thickness, and powder bed temperature) and the reward function (negative predicted porosity plus a small penalty on action magnitude to encourage feasible parameters). revision: yes

  2. Referee: [Abstract] Abstract: no information is supplied on dataset size, number of input features, train/test split, or statistical significance testing, so the claim of outperforming DQN/PPO/TD3/SAC cannot be assessed.

    Authors: We will expand the experimental section to report the dataset size, exact number of input features, train/test split ratio, and results of statistical significance tests (paired t-tests with p-values) between all compared algorithms. revision: yes

  3. Referee: [Abstract] Abstract: the assertion that multi-head attention improves exploration-exploitation balance on low-dimensional features is presented without ablation studies, attention-weight visualizations, or comparison to a non-attention SAC variant, leaving the architectural contribution unsupported.

    Authors: The manuscript already includes a direct comparison to vanilla SAC. To further substantiate the architectural contribution we will add (i) an ablation replacing multi-head attention with single-head or no attention and (ii) attention-weight heatmaps in the revised version. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical performance claims lack any derivation chain

full rationale

The manuscript reports an empirical result (convergence to 322.79 in 14 episodes, outperforming DQN/PPO/TD3/SAC) obtained by integrating multi-head attention with SAC on a porosity-prediction task. No equations, reward definitions, state representations, fitted parameters, or derivations appear in the abstract or described text. No self-citations, uniqueness theorems, or ansatzes are invoked to justify the architecture or the numerical outcome. The central claim is therefore an experimental comparison rather than a mathematical reduction, rendering the derivation chain self-contained with no load-bearing steps that collapse to their own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach relies on standard RL assumptions not detailed here.

pith-pipeline@v0.9.1-grok · 5699 in / 1009 out tokens · 32848 ms · 2026-06-26T17:07:43.272966+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Additive manufacturing for sustainable production: A survey on titanium alloys,

    P. Onu, S. M. Gad, A. A. Adediran, and C. Mbohwa, “Additive manufacturing for sustainable production: A survey on titanium alloys,” in2024 International Conference on Science, Engineering and Business for Driving Sustainable Development Goals (SEB4SDG), 2024, pp. 1–12.DOI: 10 . 1109/SEB4SDG60871.2024.10629899

    arXiv 2024

  2. [2]

    Novel sensorized additive manufacturing-based enlighted tooling concepts for aeronautical parts,

    V . Uralde, F. Veiga, A. Suarez, A. Lopez, I. Goenaga, and T. Ballesteros, “Novel sensorized additive manufacturing-based enlighted tooling concepts for aeronautical parts,”Scientific Reports, vol. 14, no. 1, p. 17 692, 2024

    2024

  3. [3]

    Aerial additive manufacturing with multiple autonomous robots,

    K. Zhang et al., “Aerial additive manufacturing with multiple autonomous robots,”Nature, vol. 609, no. 7928, pp. 709–717, 2022

    2022

  4. [4]

    Additive manufacturing and the evolution of the construction industry: A systematic lit- erature review 2015-2021,

    G. Leite and M. E. Fontana, “Additive manufacturing and the evolution of the construction industry: A systematic lit- erature review 2015-2021,” in2021 International Conference on Decision Aid Sciences and Application (DASA), 2021, pp. 294–298.DOI: 10.1109/DASA53625.2021.9682363

    work page doi:10.1109/dasa53625.2021.9682363 2015

  5. [5]

    Additive manufacturing of 3d nano- architected metals,

    A. Vyatskikh, S. Delalande, A. Kudo, X. Zhang, C. M. Portela, and J. R. Greer, “Additive manufacturing of 3d nano- architected metals,”Nature communications, vol. 9, no. 1, p. 593, 2018

    2018

  6. [6]

    Design, simulations and tests of a novel force and moments sensor for instrumented knee implants,

    P. Gasnier et al., “Design, simulations and tests of a novel force and moments sensor for instrumented knee implants,” IEEE Transactions on Biomedical Engineering, vol. 70, no. 12, pp. 3480–3489, 2023.DOI: 10 . 1109 / TBME . 2023 . 3289623

    2023

  7. [7]

    Optimizing quality inspection and con- trol in powder bed metal additive manufacturing: Challenges and research directions,

    S. Di Cataldo et al., “Optimizing quality inspection and con- trol in powder bed metal additive manufacturing: Challenges and research directions,”Proceedings of the IEEE, vol. 109, no. 4, pp. 326–346, 2021.DOI: 10.1109/JPROC.2021.3054628

    work page doi:10.1109/jproc.2021.3054628 2021

  8. [8]

    Decoding the mystery: How can LLMs turn text into Cypher in complex knowledge graphs?IEEE Access, 13:80981–81001, 2025

    B.-M. Roh, T. W. Simpson, H. Yang, S. R. T. Kumara, P. Witherell, and A. T. Jones, “Ensuring quality in metal additive manufacturing through a v-model framework,”IEEE Access, vol. 11, pp. 123 807–123 819, 2023.DOI: 10.1109/ACCESS. 2023.3327054

    work page doi:10.1109/access 2023

  9. [9]

    Data-driven pro- cess parameter optimisation for laser wire metal additive manufacturing,

    M. Roberts, M. Xia, and A. Kennedy, “Data-driven pro- cess parameter optimisation for laser wire metal additive manufacturing,” in2022 27th International Conference on Automation and Computing (ICAC), 2022, pp. 1–6.DOI: 10. 1109/ICAC55051.2022.9911139

    arXiv 2022

  10. [10]

    Physics-informed neural networks: A step towards data-driven optimization of additive manufacturing,

    F. Depaoli et al., “Physics-informed neural networks: A step towards data-driven optimization of additive manufacturing,” in2024 IEEE 29th International Conference on Emerging Technologies and Factory Automation (ETFA), 2024, pp. 1–4. DOI: 10.1109/ETFA61755.2024.10710827

    work page doi:10.1109/etfa61755.2024.10710827 2024

  11. [11]

    Cad-platform-based process optimization design method by selective laser melt- ing simulation,

    E. Dalpadulo, F. Pini, and F. Leali, “Cad-platform-based process optimization design method by selective laser melt- ing simulation,” in2020 IEEE 10th International Confer- ence Nanomaterials: Applications & Properties (NAP), 2020, 02SAMA13-1-02SAMA13–5.DOI: 10.1109/NAP51477.2020. 9309632

    work page doi:10.1109/nap51477.2020 2020

  12. [12]

    Intelligent additive manufacturing architec- ture for enhancing uniformity of surface roughness and me- chanical properties of laser powder bed fusion components,

    H.-C. Tran, Y .-L. Lo, H.-C. Yang, H.-C. Hsiao, F.-T. Cheng, and T.-H. Kuo, “Intelligent additive manufacturing architec- ture for enhancing uniformity of surface roughness and me- chanical properties of laser powder bed fusion components,” IEEE Transactions on Automation Science and Engineering, vol. 20, no. 4, pp. 2527–2538, 2023.DOI: 10.1109/TASE. 2022.3204847

    work page doi:10.1109/tase 2023

  13. [13]

    Process parameter optimization of ad- ditively manufactured parts using intelligent manufacturing,

    R. U. Rehman et al., “Process parameter optimization of ad- ditively manufactured parts using intelligent manufacturing,” Sustainability, vol. 14, no. 22, 2022,ISSN: 2071-1050.DOI: 10.3390/su142215475 [Online]. Available: https://www.mdpi. com/2071-1050/14/22/15475

    work page doi:10.3390/su142215475 2022

  14. [14]

    Machine learning to optimize additive manufacturing parameters for laser powder bed fusion of inconel 718,

    B. Kappes, S. Moorthy, D. Drake, H. Geerlings, and A. Steb- ner, “Machine learning to optimize additive manufacturing parameters for laser powder bed fusion of inconel 718,” in Proceedings of the 9th International Symposium on Super- alloy 718 & Derivatives: Energy, Aerospace, and Industrial Applications, E. Ott et al., Eds., Cham: Springer International ...

    2018

  15. [15]

    A convolutional neural network for prediction of laser power using melt-pool images in laser powder bed fusion,

    O. Kwon, H. G. Kim, W. Kim, G.-H. Kim, and K. Kim, “A convolutional neural network for prediction of laser power using melt-pool images in laser powder bed fusion,”IEEE Access, vol. 8, pp. 23 255–23 263, 2020.DOI: 10 . 1109 / ACCESS.2020.2970026

    arXiv 2020

  16. [16]

    Optimizing metal 3d printing using artificial neural network based seagull optimization algorithm,

    C. Li, “Optimizing metal 3d printing using artificial neural network based seagull optimization algorithm,” in2024 Third International Conference on Distributed Computing and Elec- trical Circuits and Electronics (ICDCECE), 2024, pp. 1–4. DOI: 10.1109/ICDCECE60827.2024.10549345

    work page doi:10.1109/icdcece60827.2024.10549345 2024

  17. [17]

    A New H I Survey of Active Galaxies

    A. J. Aghdasian, A. H. Ardakani, K. Aqabakee, and F. Ab- dollahi, “Autonomous driving using residual sensor fusion and deep reinforcement learning,” in2023 11th RSI International Conference on Robotics and Mechatronics (ICRoM), 2023, pp. 265–270.DOI: 10.1109/ICRoM60803.2023.10412516

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/icrom60803.2023.10412516 2023

  18. [18]

    Recursive generalized type-2 fuzzy radial basis function neural networks for joint position estimation and adaptive emg-based impedance control of lower limb exoskeletons,

    K. Aqabakee, F. Abdollahi, A. Taghvaeipour, and M.-R. Akbarzadeh-T, “Recursive generalized type-2 fuzzy radial basis function neural networks for joint position estimation and adaptive emg-based impedance control of lower limb exoskeletons,”Biomedical Signal Processing and Control, vol. 100, p. 106 791, 2025,ISSN: 1746-8094.DOI: https : / / doi . org / 10...

    2025

  19. [19]

    A fast balance op- timization approach for charging enhancement of lithium-ion battery packs through deep reinforcement learning,

    A. Heydarian Ardakani and F. Abdollahi, “A fast balance op- timization approach for charging enhancement of lithium-ion battery packs through deep reinforcement learning,”Journal of Energy Storage, vol. 89, p. 111 755, 2024,ISSN: 2352-152X. DOI: https://doi.org/10.1016/j.est.2024.111755 [Online]. Available: https://www.sciencedirect.com/science/article/pi...

    work page doi:10.1016/j.est.2024.111755 2024

  20. [20]

    Scalable charging optimization of battery energy storage systems with deep reinforcement learning,

    A. H. Ardakani, K. Aqabakee, F. Abdollahi, and E. Shirazi, “Scalable charging optimization of battery energy storage systems with deep reinforcement learning,” in2024 IEEE PES Innovative Smart Grid Technologies Europe (ISGT EUROPE), 2024, pp. 1–5.DOI: 10 . 1109 / ISGTEUROPE62998 . 2024 . 10863295

    2024

  21. [21]

    A novel porosity prediction frame- work based on reinforcement learning for process parameter optimization in additive manufacturing,

    A. M. Faizan Mohamed, F. Careri, R. H. Khan, M. M. Attallah, and L. Stella, “A novel porosity prediction frame- work based on reinforcement learning for process parameter optimization in additive manufacturing,”Scripta Materialia, vol. 255, p. 116 377, 2025,ISSN: 1359-6462.DOI: https : / / doi . org / 10 . 1016 / j . scriptamat . 2024 . 116377 [Online]. A...

    2025

  22. [22]

    A reinforce- ment learning approach for process parameter optimization in additive manufacturing,

    S. Dharmadhikari, N. Menon, and A. Basak, “A reinforce- ment learning approach for process parameter optimization in additive manufacturing,”Additive Manufacturing, vol. 71, p. 103 556, 2023,ISSN: 2214-8604.DOI: https://doi.org/10. 1016/j.addma.2023.103556 [Online]. Available: https://www. sciencedirect.com/science/article/pii/S2214860423001690

    arXiv 2023

  23. [23]

    Reinforcement learning and optimization based path planning for thin-walled structures in wire arc additive manufacturing,

    J. Petrik and M. Bambach, “Reinforcement learning and optimization based path planning for thin-walled structures in wire arc additive manufacturing,”Journal of Manufacturing Processes, vol. 93, pp. 75–89, 2023,ISSN: 1526-6125.DOI: https : / / doi . org / 10 . 1016 / j . jmapro . 2023 . 03 . 013 [Online]. Available: https://www.sciencedirect.com/science/a...

    2023

  24. [24]

    Application of machine learning in additive manufacturing of a novel al alloy heat exchanger,

    F. Careri, L. Stella, R. H. U. Khan, and M. M. Attallah, “Application of machine learning in additive manufacturing of a novel al alloy heat exchanger,” 2025.DOI: 10.1007/s00170- 025-15389-y

    work page doi:10.1007/s00170- 2025

  25. [25]

    M. Buhairi et al.,Review on volumetric energy density: Influence on morphology and mechanical properties of ti6al4v manufactured via laser powder bed fusion, prog addit manuf 8 (2023) 265–283

    2023

  26. [26]

    Human-level control through deep reinforce- ment learning,

    V . Mnih et al., “Human-level control through deep reinforce- ment learning,”nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  27. [27]

    Addressing function approximation error in actor-critic methods,

    S. Fujimoto, H. van Hoof, and D. Meger, “Addressing function approximation error in actor-critic methods,”CoRR, vol. abs/1802.09477, 2018. arXiv: 1802 . 09477. [Online]. Available: http://arxiv.org/abs/1802.09477

    Pith/arXiv arXiv 2018

  28. [28]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”CoRR, vol. abs/1707.06347, 2017. arXiv: 1707 . 06347. [Online]. Available: http://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  29. [29]

    Soft actor- critic: Off-policy maximum entropy deep reinforcement learn- ing with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor- critic: Off-policy maximum entropy deep reinforcement learn- ing with a stochastic actor,”CoRR, vol. abs/1801.01290, 2018. arXiv: 1801.01290. [Online]. Available: http://arxiv.org/abs/ 1801.01290

    Pith/arXiv arXiv 2018

  30. [30]

    Tensor2tensor for neural machine transla- tion,

    A. Vaswani et al., “Tensor2tensor for neural machine transla- tion,”CoRR, vol. abs/1803.07416, 2018. [Online]. Available: http://arxiv.org/abs/1803.07416

    Pith/arXiv arXiv 2018

  31. [31]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”Advances in Neural Information Processing Systems, 2017

    2017

  32. [32]

    Zhang, L

    T. Zhang, L. Stella, and J. Barreiro-Gomez,Experience-replay innovative dynamics, 2025. arXiv: 2501 . 12199[cs.LG]. [Online]. Available: https://arxiv.org/abs/2501.12199

    arXiv 2025

  33. [33]

    A multi- agent reinforcement learning approach to promote cooperation in evolutionary games on networks with environmental feed- back,

    T. Zhang, H. Gupta, K. Suprabhat, and L. Stella, “A multi- agent reinforcement learning approach to promote cooperation in evolutionary games on networks with environmental feed- back,” in2023 62nd IEEE Conference on Decision and Con- trol (CDC), 2023, pp. 2196–2201.DOI: 10.1109/CDC49753. 2023.10383787 Kianoush Aqabakeereceived the B.S. degree in Electric...

    work page doi:10.1109/cdc49753 2023

  34. [34]

    His research interests include computational intelligence, control theory, and biomedical engineering

    Since 2025, he has been a PhD student in the Department of Information Technology, Faculty of Engineering and Architecture, Ghent University, Belgium. His research interests include computational intelligence, control theory, and biomedical engineering. Leonardo Stellareceived the Laurea Triennale de- gree in Computer Engineering in 2013 from the Universi...

    2025