Pith. sign in

REVIEW 5 major objections 4 minor 33 references

Reinforcement Learning on Reconfigurable Hardware: Overcoming Material Variability in Laser Material Processing

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An RL policy implemented on an FPGA learns per-surface laser power profiles that beat the best constant power by up to 23%.

desk verdict A genuinely new real-time RL laser-welding controller on FPGA, but the headline gains are computed in the same reward signal the agent optimizes; the engineering is real, the quantitative claim needs independent weld metrics. read the letter →

arxiv 2501.19102 v2 pith:UVTJDIHQ submitted 2025-01-31 cs.LG

Giulio Masinelli , Chang Rajani , Patrik Hoffmann , Kilian Wasmer , David Atienza This is my paper · ORCID
classification cs.LG
keywords reinforcementlearninglaserweldingFPGAreal-timecontrolsoftactor-criticsurfaceroughnessadaptationopticalreflectionmaterialprocessing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that reinforcement learning can control a laser welder in real time, compensating for changing surface roughness without a predetermined target or per-surface calibration. It mounts a small neural-network policy on an FPGA that reads optical signals and sets laser power every 10 ms, while a server retrains the policy between welds using soft actor-critic and a reward that is simply the scaled optical reflection signal. On 316L stainless steel with brushed, sandblasted, and mixed surfaces, the learned policy outperforms the best constant-power setting by about 23% on rougher surfaces and 7% on mixed ones, and produces larger, deeper melt pools in cross-section. If this holds, it points to a practical way to automate laser processes that currently require manual or simulation-based tuning.

What carries the argument

The central object is an FPGA-resident multilayer perceptron implementing a soft actor-critic policy, a maximum-entropy RL algorithm that learns a stochastic policy balancing reward with exploration. The reward is $r(s_{t+1}) = OR(s_{t+1})/10$, where $OR$ is the on-axis optical reflection from the process zone; the paper adopts the empirical finding that $OR$ peaks just before the melt pool leaves conduction mode, corresponding to the largest stable weld. Inference takes 3.54 microseconds within each 10 ms control period, while training runs on a server between episodes using quantized weights that are streamed back to the FPGA, so the policy can react to surface changes inside a single weld line.

What would settle it

Weld sandblasted 316L samples at a range of constant powers and find where measured penetration depth is largest without keyhole porosity; then compare that power with the power the learned policy chooses and with the OR peak. If the OR peak occurs at a power that does not produce the largest stable weld, the reward assumption fails, and a policy trained on $r = OR/10$ would be optimizing the wrong quantity.

Watch

Extended reading notes

Core claim

The paper claims that a closed-loop RL controller for laser welding can run at microsecond latency on an FPGA and adapt laser power to the surface it is welding with no per-surface calibration. The policy is a soft actor-critic network taking optical reflection and emission signals as observations and outputting laser power every 10 ms, trained between episodes on a server with the reward $r(s_{t+1}) = OR(s_{t+1})/10$. This reward encodes the finding that the optical reflection signal peaks just before the melt pool transitions from conduction mode to keyhole mode, marking the largest stable weld. In laser welding tests on 316L stainless steel, the learned policy achieved test returns 6.8% higher on brushed, 22.6% higher on sandblasted, and 7.3% higher on mixed surfaces than the best constant-power baseline, and post-weld cross-sections showed wider and deeper melt pools without keyhole formation.

Load-bearing premise

The load-bearing premise is that the optical reflection signal is a trustworthy measure of weld quality, peaking exactly at the largest stable weld before keyhole formation, so a policy maximizing $r = OR/10$ is actually maximizing good welds.

Editorial extensions

If this is right

  • FPGA inference at 3.54 microseconds per action means the control loop can respond to surface changes within a single weld line, not just between samples.
  • No predetermined weld depth or bead width target is required; the same reward and learning loop adapt across brushed, sandblasted, and mixed surfaces.
  • Dynamic power profiles can beat an optimized constant-power baseline, with test returns 6.8%, 22.6%, and 7.3% higher on the three surface types.
  • The learned policy keeps the weld in conduction mode and avoids keyhole formation, and post-weld cross-sections show wider and deeper melt pools associated with better penetration and fusion.
  • Because the reward can in principle be any sensor signal, the server-train/FPGA-infer architecture is a general template for other laser processes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test the paper does not run is mechanical testing: cutting and tensile-testing welds made by the learned policy versus constant power would settle whether the OR-based gains translate into strength, since only cross-section geometry is reported.
  • If the OR-to-quality relation holds beyond 316L steel, the same reward and control loop could be ported to other materials and laser processes, but that generalization is extrapolation from a single alloy.
  • The policy's observed behavior on mixed surfaces, with a power spike at ignition and raised power in sandblasted regions, suggests it learned a physically interpretable control law; a useful next probe is whether the policy transfers to unseen surface roughness values or retrains from scratch.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper presents a closed-loop laser welding control system in which a Soft Actor-Critic policy, trained on a server with quantization-aware training and deployed on an FPGA, adjusts laser power in real time (3.54 microsecond inference) based on optical reflection (OR) and emission (OE) signals. The reward is r(st+1)=OR(st+1)/10, motivated by the claim that the OR signal peaks just before keyhole transition and corresponds to the largest stable conduction-mode weld. Experiments on brushed, sandblasted, and mixed 316L stainless steel surfaces compare the learned policy against an optimal constant power baseline selected by grid search on average OR. The paper reports 6.8%, 22.58%, and 7.30% improvements in test return for the three surface types, and shows representative cross-sections suggesting larger melt pools for the learned policy.

Significance. If the quantitative claims are supported, this would be a notable engineering contribution: the first demonstration of a closed-loop RL controller for laser welding on reconfigurable hardware at microsecond latency, with no predetermined target and with real-world surface variability. The paper's strengths include a fully described hardware-in-the-loop system, explicit pseudocode, reproducible hyperparameters, real experiments rather than simulation-only validation, and qualitative figures showing adaptive power profiles. However, the central quantitative result is measured in the same OR signal that defines both the reward and the baseline selection, and no independent physical weld-quality metrics are quantified. The significance of the work as a 'better welding' claim therefore depends on additional validation that is not currently in the manuscript.

major comments (5)
  1. [IV, Eq. (1)] The reported improvements (6.8%, 22.58%, 7.30% in Section IV) are computed on the same reward r(st+1)=OR(st+1)/10 (Eq. (1)) that defines the RL objective, and the optimal-constant-power baseline is itself selected by grid search for the highest average OR signal (Section IV). The comparison therefore demonstrates that the learned policy achieves higher OR-based returns than the best constant-power OR policy, but it does not by itself establish better weld quality. The manuscript's own Discussion (Section V) and Conclusion (Section VI) concede that 'the ultimate quality of the weld is determined by physical metrics,' yet no quantitative physical metric is reported: no melt-pool depth/width measurements, porosity counts, or mechanical tests, and no sample-size or error-bar information for the cross-sections in Fig. 4. I request quantitative, independent weld-quality metrics (e.g., melt-pool depth and width from multiple cross-sections, porosity analysis, or tensile tests) with means and variances for both policies. Without these, a policy that raises OR by exploiting sensor dynamics or plume reflections would also be reported as 'better.'
  2. [IV, Fig. 2] The headline percentages are given without any measure of variability. The test-episode return plots in Fig. 2 show substantial run-to-run variation, especially for the mixed-surface condition, and the number of test episodes used to compute the reported improvements is not stated. Please report the mean and standard deviation (or confidence interval) over at least N independent test episodes per surface condition, and state N, for both the learned policy and the constant-power baseline. This is necessary to determine whether the claimed 6.8–22.6% improvements are statistically significant.
  3. [III.A, Eq. (1)] The paper claims the method works 'without relying on reward engineering' (Abstract and Section I), but the reward is explicitly chosen based on the physics finding of Wittemer et al. [19] that the OR signal peaks just before keyhole transition, and Eq. (1) scales OR by 1/10. This is a reward engineered from domain knowledge, not a reward-free formulation. The statement in Section III.A that 'we present a system that can in principle be used with any reward' is a reasonable framing, but the abstract and introduction should be revised to avoid overclaiming; the contribution is a specific domain-informed reward proxy.
  4. [III.A and Algorithm pseudocode] The state observation is never precisely defined. The text says observations are 'derived from two optical sensors' and the pseudocode says 'Acquire optical data st', but the exact construction of st (e.g., raw samples of OR and OE over a 10 ms window, downsampled/aggregated statistics, normalization) is not specified. This omission prevents reproduction and makes it impossible to assess the POMDP formulation. Please specify the observation vector dimensions and preprocessing.
  5. [III.B] The Brevitas digital twin is used to simulate FPGA quantization during QAT, but no validation is shown that the deployed 8-bit integer policy on the FPGA produces actions consistent with the floating-point SAC policy trained on the server. Given that the central claim is closed-loop control on the FPGA, please report an action-distribution comparison (or a deployment-vs-digital-twin return comparison) to quantify the quantization gap.
minor comments (4)
  1. [IV] Please state the grid-search resolution for the constant-power baseline and the resulting optimal power values for each surface type; the current text only says 'grid search over the laser power range from 25W to 100W.'
  2. [Fig. 2] The red dashed line for 'Optimal Constant Power' should clarify whether it is the return of a single constant-power episode or the mean over multiple episodes; a single line without error bars or a sample-size statement is ambiguous.
  3. [Fig. 3] The step numbers in the text (e.g., 'steps 20 to 60') correspond to 10 ms steps at 50 mm/s, i.e., 10–30 mm, which matches the sandblasted segment of the mixed sample; stating this correspondence explicitly would help the reader interpret the adaptive power profile.
  4. [Abstract] The abstract states the method works 'without relying on reward engineering or prior setup information'; please align this phrasing with the reward discussion in Section III.A, which acknowledges the reward choice is based on prior physics knowledge from [19].

Circularity Check

1 steps flagged · score 4.0 of 10

Quantitative gains are measured in the same OR signal used as the reward and as the baseline-selection criterion, so the headline improvement partly reduces to the training objective; weld-quality conclusions rely on an uncalibrated proxy.

  1. fitted input called prediction [Section III-A, Eq. (1); Section IV (optimal constant power selection); Conclusion]
    "we choose a simple reward function which encourages large stable welds while staying within conduction mode (avoiding vaporization of material), using the OR signal as the reward: r(st+1) = OR(st+1)/10. ... Specifically, we selected the power setting that yielded the highest average OR signal along the weld lines, correlating with the best weld quality in conduction mode achievable with traditional constant power welding. ... ultimately achieving a 7.30% improvement in terms of test return."

    The reported improvements (6.8%, 22.58%, 7.30%) are episode returns, i.e., cumulative values of the same r=OR/10 that the SAC agent is trained to maximize, and the optimal constant power baseline is itself selected by grid search to maximize average OR on the same weld lines. The quantitative claim that RL beats optimal constant power is therefore a comparison of two OR-maximizing policies on the OR objective; it does not independently validate the paper's stronger conclusion of improved weld quality. The paper's own Conclusion concedes that 'the ultimate quality of the weld is determined by physical metrics,' and the only non-OR evidence is qualitative cross-section images without measured melt-pool dimensions, porosity counts, or error bars.

full rationale

The FPGA/RL implementation itself is self-contained: the learned policy, quantization-aware training, and real-time inference do not reduce to the inputs. No load-bearing self-citation or imported uniqueness theorem appears; citation [8] is prior work by the same authors but only contextual. The main circularity concern is evaluative. Eq. (1) defines the reward as OR/10, and Section IV defines the baseline as the constant power maximizing average OR. Because the reported percentage improvements are computed from episode returns of that same reward, the headline numbers partly restate that the trained policy achieves higher scores on the function it was trained to optimize. Whether that implies better welds depends entirely on the Wittemer et al. proxy that OR peaks at the largest stable conduction-mode weld. The paper is transparent about this limitation in the Discussion and Conclusion, and the qualitative cross-sections provide some independent but non-quantified support. For these reasons the circularity is partial rather than total: the control result is meaningful, but the central quality claim is not independently established beyond the reward signal.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are postulated. The only additions are engineering components (FPGA in-the-loop control, digital twin) and hand-chosen constants, which are covered under free parameters and axioms. The load-bearing external input is the cited OR-to-weld-quality relationship, which the paper imports as an axiom.

free parameters (3)
  • Optimal constant power per surface (baseline) = not reported (grid-searched over 25-100 W)
    The baseline used to compute the claimed 6.8%, 22.6%, and 7.3% improvements is selected per surface type by maximizing mean OR reward. Its exact values are not stated, so the margin depends on an unreported fitted quantity.
  • OR reward scaling factor (1/10) = 0.1
    Hand-chosen normalization to map 0-10 V photodiode output to a 0-1 reward. It scales all returns but does not change the ranking of policies; included for completeness.
  • OR acquisition trigger threshold = 0.1 V
    The pseudocode only acquires and controls while OR >= 0.1 V. This hand-chosen threshold affects when the controller is active; its impact is not analyzed.
assumptions (3)
  • domain assumption OR signal peaks just before the conduction-to-keyhole transition and corresponds to the largest stable weld [19]
    The entire reward design (Eq. 1) rests on this cited relationship. If OR is not a monotone proxy for weld size/quality in conduction mode, maximizing OR may not improve weld quality.
  • domain assumption The optical reflection and emission signals provide sufficient information to choose the correct laser power
    The POMDP observation ot is fed to a memoryless MLP; the paper assumes the two photodiode signals resolve the relevant melt-pool state. The authors themselves note perceptual aliasing as a limitation in the Discussion.
  • ad hoc to paper The Brevitas digital twin accurately models FPGA quantization for QAT-trained weights
    The server trains with a simulated FPGA and streams weights to the PL BRAM; the paper states a scaling factor ensures consistency but does not quantify the residual quantization error between twin and hardware.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement Learning on Reconfigurable Hardware: Overcoming Material Variability in Laser Material Processing." pith.science (2026). https://pith.science/paper/UVTJDIHQ

@misc{pith2026250119102,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning on Reconfigurable Hardware: Overcoming Material Variability in Laser Material Processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVTJDIHQ}},
  note         = {Machine review of arXiv:2501.19102}
}
read the original abstract

Ensuring consistent processing quality is challenging in laser processes due to varying material properties and surface conditions. Although some approaches have shown promise in solving this problem via automation, they often rely on predetermined targets or are limited to simulated environments. To address these shortcomings, we propose a novel real-time reinforcement learning approach for laser process control, implemented on a Field Programmable Gate Array to achieve real-time execution. Our experimental results from laser welding tests on stainless steel samples with a range of surface roughnesses validated the method's ability to adapt autonomously, without relying on reward engineering or prior setup information. Specifically, the algorithm learned the correct power profile for each unique surface characteristic, demonstrating significant improvements over hand-engineered optimal constant power strategies -- up to 23% better performance on rougher surfaces and 7% on mixed surfaces. This approach represents a significant advancement in automating and optimizing laser processes, with potential applications across multiple industries.

Figures

Figures reproduced from arXiv: 2501.19102 by the authors.

Figure 1
Figure 1. Illustration of the proposed method. Top: The FPGA receives optical signals from the process zone and uses its onboard policy network to determine the laser power in real-time. Between processing runs, the collected data is sent to a server where RL is used to train the policy. Bottom: The policy initially starts with random actions and learns to optimize the process outcome, achieving the best possible results whil… view at source ↗
Figure 2
Figure 2. Comparison of our RL algorithm performance on different sample [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of OR and laser power actions during a test episode. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Representative melt-pool cross-sections taken from the mixed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 25 canonical work pages

  1. [19]

    Absorbance measurement for in situ process regime identification in laser processing,

    M. Wittemer, J. Gr ¨unewald, and K. Wudy, “Absorbance measurement for in situ process regime identification in laser processing,” The International Journal of Advanced Manufacturing Technology , vol. 126, no. 1, pp. 103–115, May 2023. [Online]. Available: https://doi.org/10.1007/s00170-023-11041-9

    work page doi:10.1007/s00170-023-11041-9 2023

  2. [1]

    Laser-based manufacturing concepts for efficient production of tailor welded sheet metals,

    M. Sp ¨ottl and H. Mohrbacher, “Laser-based manufacturing concepts for efficient production of tailor welded sheet metals,” Advances in Manufacturing , vol. 2, no. 3, pp. 193–202, sep

    work page

  3. [2]

    Laser powder bed fusion of 316l stainless steel: Industrial relevance and process optimization,

    I. Serrano-Munoz, J.-Y . Buffiere, R. Mokso, C. Verdu, and Y . Nadot, “Laser powder bed fusion of 316l stainless steel: Industrial relevance and process optimization,” Advanced Engineering Materials , vol. 22, no. 3, p. 1900617, 2020

    work page 2020

  4. [3]

    A review on laser processing in electronic and mems packaging,

    S. Liu, Y . Chen, H. Guo, X. Tian, H. Lu, J. Zhao, and G. Zhang, “A review on laser processing in electronic and mems packaging,” Journal of Electronic Packaging , vol. 139, no. 3, p. 030801, 9 2017. [Online]. Available: https://doi.org/10.1115/1.4036605

    work page doi:10.1115/1.4036605 2017

  5. [4]

    W. M. Steen and J. Mazumder, Laser Material Processing , 4th ed. Springer London, 2010. [Online]. Available: https: //doi.org/10.1007/978-1-84996-062-5

    work page doi:10.1007/978-1-84996-062-5 2010

  6. [5]

    Fault-tolerant laser welding in automotive industries,

    M. Schmidt, A. Otto, A. Grimm, and C. K ¨ageler, “Fault-tolerant laser welding in automotive industries,” ICALEO 2006 - 25th International Congress on Applications of Laser and Electro-Optics, Congress Proceedings, oct 2006. [Online]. Available: /lia/liacp/article/doi/10. 2351/1.5060833/834817/Fault-tolerant-laser-welding-in-automotive

    work page 2006

  7. [6]

    Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,

    S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,” in 2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017, pp. 3389–3396

    2017

  8. [7]

    Magnetic control of tokamak plasmas through deep reinforcement learning,

    J. Degrave, F. Felici, J. Buchli, M. Neunert, B. Tracey, F. Carpanese, T. Ewalds, R. Hafner, A. Abdolmaleki, D. de Las Casas, et al. , “Magnetic control of tokamak plasmas through deep reinforcement learning,” Nature, vol. 602, no. 7897, pp. 414–419, 2022

    work page 2022

Show all 33 references
  1. [8]

    Adaptive laser welding control: A reinforcement learning approach,

    G. Masinelli, T. Le-Quang, S. Zanoli, K. Wasmer, and S. A. Shevchik, “Adaptive laser welding control: A reinforcement learning approach,” Ieee Access, vol. 8, pp. 103 803–103 814, 2020

    2020

  2. [9]

    Reinforcement learning for laser welding speed control minimizing bead width error,

    T. Kaneko, G. Minamoto, Y . Hirose, and T. Sakai, “Reinforcement learning for laser welding speed control minimizing bead width error,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 12 275–12 281

    2023

  3. [10]

    In- telligent laser welding through representation, prediction, and control learning: An architecture with deep neural networks and reinforcement learning,

    J. G ¨unther, P. M. Pilarski, G. Helfrich, H. Shen, and K. Diepold, “In- telligent laser welding through representation, prediction, and control learning: An architecture with deep neural networks and reinforcement learning,” Mechatronics, vol. 34, pp. 1–11, 2016

    2016

  4. [11]

    Development of a defect-detection platform using photodiode signals collected from the melt pool of laser powder-bed fusion,

    K. Taherkhani, E. Sheydaeian, C. Eischer, M. Otto, and E. Toyserkani, “Development of a defect-detection platform using photodiode signals collected from the melt pool of laser powder-bed fusion,” Additive Manufacturing, vol. 46, p. 102152, 10 2021

    2021

  5. [12]

    Steel corrosion behaviour in acidic solution for application in petrochemical distillation systems,

    M. A. Fajobi, R. T. Loto, and O. O. Oluwole, “Steel corrosion behaviour in acidic solution for application in petrochemical distillation systems,” IOP Conference Series: Materials Science and Engineering, vol. 811, no. 1, p. 012030, apr 2020. [Online]. Available: https://dx.do...

    2020 doi

  6. [13]

    Anti-fouling surface modified stainless steel for food processing,

    J. A. Barish and J. M. Goddard, “Anti-fouling surface modified stainless steel for food processing,” Food and Bioproducts Processing, vol. 91, no. 4, pp. 352–361, 2013. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0960308513000047

    2013

  7. [14]

    Electrophoretic deposition of natural hydroxyapatite on medical grade 316l stainless steel,

    M. Javidi, S. Javadpour, M. Bahrololoom, and J. Ma, “Electrophoretic deposition of natural hydroxyapatite on medical grade 316l stainless steel,” Materials Science and Engineering: C , vol. 28, no. 8, pp. 1509–1515, 2008. [Online]. Available: https://www.sciencedirect.com/ sci...

    2008

  8. [15]

    Laser beam welding of 316l t-joint: microstructure, microhardness, distortion, and residual stress,

    Y . Rong, Y . Huang, G. Zhang, G. Mi, and W. Shao, “Laser beam welding of 316l t-joint: microstructure, microhardness, distortion, and residual stress,” The International Journal of Advanced Manufacturing Technology, vol. 90, no. 5, pp. 2263–2270, May 2017. [Online]. Available...

    2017 doi

  9. [16]

    Xilinx/brevitas,

    A. Pappalardo, “Xilinx/brevitas,” 2023. [Online]. Available: https: //doi.org/10.5281/zenodo.3333552

    2023 doi

  10. [17]

    Reinforcement learning: An introduction,

    R. S. Sutton, “Reinforcement learning: An introduction,” A Bradford Book, 2018

    2018

  11. [18]

    Unpack- ing reward shaping: Understanding the benefits of reward engineering on sample complexity,

    A. Gupta, A. Pacchiano, Y . Zhai, S. Kakade, and S. Levine, “Unpack- ing reward shaping: Understanding the benefits of reward engineering on sample complexity,” Advances in Neural Information Processing Systems, vol. 35, pp. 15 281–15 295, 2022

    2022

  12. [20]

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

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” in International conference on machine learning . PMLR, 2018, pp. 1861–1870

    2018

  13. [21]

    Soft actor-critic algorithms and applications,

    T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V . Kumar, H. Zhu, A. Gupta, P. Abbeel, et al. , “Soft actor-critic algorithms and applications,” arXiv preprint arXiv:1812.05905, 2018

    2018 arXiv

  14. [22]

    Maximum entropy rl (provably) solves some robust rl problems,

    B. Eysenbach and S. Levine, “Maximum entropy rl (provably) solves some robust rl problems,” arXiv preprint arXiv:2103.06257 , 2021

    2021 arXiv

  15. [23]

    [Online]

    Xilinx, Introduction to FPGA Design with Vivado High-Level Synthesis, Xilinx Inc., 2019, uG998 (v1.2). [Online]. Available: https: //docs.xilinx.com/v/u/en-US/ug998-vivado-intro-fpga-design-hls

    2019

  16. [24]

    Auto-encoding variational bayes,

    D. P. Kingma, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  17. [25]

    Keyhole fluctuation and pore formation mechanisms during laser powder bed fusion additive manufacturing,

    Y . Huang, T. G. Fleming, S. J. Clark, S. Marussi, K. Fezzaa, J. Thiyagalingam, C. L. A. Leung, and P. D. Lee, “Keyhole fluctuation and pore formation mechanisms during laser powder bed fusion additive manufacturing,” Nature Communications , vol. 13, no. 1, p. 1170, 2022. [Onl...

    2022

  18. [26]

    A review on melt-pool characteristics in laser welding of metals,

    B. Fotovvati, S. F. Wayne, G. Lewis, and E. Asadi, “A review on melt-pool characteristics in laser welding of metals,” Advances in Materials Science and Engineering , vol. 2018, no. 1, p. 4920718,

    2018

  19. [27]

    Control of meltpool shape in laser welding,

    W. Suder, X. Chen, D. R. Sierra, G. Chen, J. Wainwright, K. Rajamudili, G. R. Pardal, and S. Williams, “Control of meltpool shape in laser welding,” Welding in the World , vol. 68, no. 6, pp. 1485–1495, June 2024. [Online]. Available: https: //doi.org/10.1007/s40194-024-01719-3

    2024 doi

  20. [28]

    Defectometry, distortion analysis and metallurgical properties in the keyhole and conduction mode of laser beam welding of al–mg–sc alloy butt joints,

    K. Barat, C. N. Kumar, G. Patil, K. Panbarasu, and K. Venkateswarlu, “Defectometry, distortion analysis and metallurgical properties in the keyhole and conduction mode of laser beam welding of al–mg–sc alloy butt joints,” Optics & Laser Technology , vol. 162, p. 109248, 2023. ...

    2023

  21. [29]

    Effect of laser beam welding on microstructure, tensile strength and fatigue behaviour of duplex stainless steel 2205,

    A. E. Odermatt, V . Ventzke, F. Dorn, R. Dins ´e, P. Merhof, and N. Kashaev, “Effect of laser beam welding on microstructure, tensile strength and fatigue behaviour of duplex stainless steel 2205,” Journal of Manufacturing Processes , vol. 72, pp. 148–158, 2021. [Online]. Avai...

    2021

  22. [30]

    Reinforcement learning with perceptual aliasing: The perceptual distinctions approach,

    L. Chrisman, “Reinforcement learning with perceptual aliasing: The perceptual distinctions approach,” in Proceedings of the AAAI Conference on Artificial Intelligence , 1992. [Online]. Available: https://api.semanticscholar.org/CorpusID:1963904

    1992

  23. [31]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in 2017 IEEE/RSJ international con- ference on intelligent robots and systems (IROS) . IEEE, 2017, pp. 23–30

    2017

  24. [2014]

    Available: https://link.springer.com/article/10.1007/ s40436-014-0088-8

    [Online]. Available: https://link.springer.com/article/10.1007/ s40436-014-0088-8

  25. [2018]

    Available: https://onlinelibrary.wiley.com/doi/abs/10

    [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10. 1155/2018/4920718

    2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.