pith. sign in

arxiv: 2606.10227 · v1 · pith:SU32ZYNTnew · submitted 2026-06-08 · 💻 cs.LG

Spatiotemporal Graph Transformer for 3D Neighborhood Interaction and Quality Prediction in Metal Additive Manufacturing

Joyce Karen Pelaez , Siqi Zhang , Hoo Sang Ko This is my paper

Pith reviewed 2026-06-27 16:59 UTC · model grok-4.3

classification 💻 cs.LG
keywords spatiotemporal graph transformermetal additive manufacturingquality prediction3D neighborhood interactionsweighted network representationdual-attention transformercross-layer interactionsmultimodal data integration
0
0 comments X

The pith

A dual-attention spatiotemporal graph transformer on weighted networks models 3D neighborhood interactions to predict build quality in metal additive manufacturing.

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

The paper develops a weighted network representation that treats fusing locations as nodes and encodes their spatial- and process-dependent relationships as edge weights. This structure unifies geometric design, process settings, and in-situ sensing data for downstream tasks. A dual-attention graph transformer is then applied to capture both within-node feature dependencies and cross-node neighborhood interactions across layers. Experiments show the approach outperforms image-based, sequence-based, and graph-based models, with the explicit modeling of cross-layer interactions proving essential for accurate quality prediction.

Core claim

A weighted network representation of the manufacturing process models fusing locations as nodes whose spatial- and process-dependent relationships appear as edge weights; this representation also integrates multimodal data into one structure. A dual-attention graph transformer built on the network learns within-node feature dependencies together with cross-node neighborhood interactions. The resulting quality representations yield significantly higher prediction accuracy than image-based, sequence-based, or standard graph-based alternatives, and ablation confirms that cross-layer interactions are critical to the gains.

What carries the argument

Weighted network representation of fusing locations as nodes with spatial- and process-dependent edge weights, which feeds a dual-attention graph transformer that jointly models within-node features and cross-node 3D neighborhood interactions.

If this is right

  • Incorporating cross-layer interactions measurably improves quality-prediction accuracy over models that ignore them.
  • The network structure unifies geometric, process, and sensing inputs into a single learnable representation.
  • The dual-attention mechanism extracts both local feature dependencies and neighborhood effects that prior image and sequence models miss.
  • The same network-plus-transformer pattern applies to other tasks that require modeling network-structured manufacturing data.

Where Pith is reading between the lines

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

  • The same node-and-edge construction could be tested on other layer-wise processes such as directed-energy deposition to check whether cross-layer effects remain dominant.
  • Closed-loop parameter adjustment might become feasible if the learned quality representations are fed back into process controllers in real time.
  • Adding explicit temporal attention layers could reveal whether quality effects accumulate over many build cycles rather than only adjacent layers.

Load-bearing premise

The weighted network representation of fusing locations with spatial- and process-dependent edge weights accurately captures 3D neighborhood interactions and supports effective multimodal integration.

What would settle it

If an ablation that removes the cross-layer interaction terms from the dual-attention transformer shows no drop in quality-prediction accuracy on the same experimental datasets, the claim that those interactions are critical would be falsified.

read the original abstract

Metal additive manufacturing enables the fabrication of complex parts, but achieving consistent build quality remains challenging due to interactions induced by repeated layer-wise melting, solidification, and reheating across the 3D build. Advanced sensing provide a great opportunity to collect rich observations of the actual manufacturing process for real-time quality monitoring and control. Yet, existing methods often have limited ability to represent multi-layer interactions and quantify their contributions to quality. In this paper, we develop a novel spatiotemporal graph transformer for modeling 3D neighborhood interactions and learn their effects on build quality in metal additive manufacturing. Specifically, we first introduce a weighted network representation of the manufacturing process, where fusing locations are modeled as nodes, and their spatial- and process-dependent relationships are encoded as edge weights. This representation also enables the integration of multimodal data (e.g., geometric design, process settings, and in-situ sensing data) into a unified structure for downstream learning tasks. Building on this network, we further design a dual-attention graph transformer that captures both within-node feature dependencies and cross-node neighborhood interactions for quality representation learning. Experimental results show that the proposed framework significantly outperforms image-based, sequence-based, and graph-based models in characterizing process-quality relationships. More importantly, the incorporation of cross-layer interactions is critical for improving quality prediction performance. This framework is broadly applicable to other tasks involving network modeling and graph-based representation learning.

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

0 major / 2 minor

Summary. The manuscript introduces a spatiotemporal graph transformer for modeling 3D neighborhood interactions and predicting build quality in metal additive manufacturing. It first defines a weighted network representation in which fusing locations are nodes and spatial- and process-dependent relationships are encoded as edge weights; this structure also fuses multimodal inputs (geometric design, process settings, in-situ sensing). A dual-attention graph transformer is then applied to capture within-node feature dependencies and cross-node (including cross-layer) interactions. The central empirical claim is that the resulting framework significantly outperforms image-based, sequence-based, and graph-based baselines, and that the explicit modeling of cross-layer interactions is critical to performance.

Significance. If the experimental claims are substantiated with appropriate datasets, baselines, and statistical controls, the work would offer a concrete graph-construction method for integrating heterogeneous AM sensing data and a dual-attention mechanism that explicitly targets 3D neighborhood effects. The emphasis on cross-layer interactions addresses a recognized gap in current layer-wise monitoring approaches.

minor comments (2)
  1. The abstract states outperformance and the importance of cross-layer terms but supplies no information on datasets, metrics, statistical tests, baseline implementations, or potential confounds; the experimental section should include these details to allow assessment of the central claim.
  2. Notation for the weighted network (node features, edge-weight definitions) and the dual-attention mechanism should be introduced with explicit equations and a small illustrative diagram early in §3 to improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their summary of our work and for noting its potential significance in addressing cross-layer interactions via graph-based modeling in metal additive manufacturing. No specific major comments were listed in the report, so we provide no point-by-point responses below. We are happy to address any additional questions or concerns the referee may have.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper presents a new weighted network representation of fusing locations with spatial- and process-dependent edge weights, followed by a dual-attention graph transformer to capture within-node and cross-node interactions for quality prediction. The central claims are supported by experimental outperformance against baselines and the stated importance of cross-layer terms. No equations, derivations, or self-citations in the abstract or reader's summary reduce any prediction or uniqueness claim to fitted inputs or prior self-referential results by construction. The model is introduced as an independent construction, with performance evaluated externally via comparisons, making the derivation self-contained against benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on a new graph-based representation and attention mechanism introduced without external benchmarks or independent evidence in the abstract; standard graph assumptions are invoked implicitly.

axioms (1)
  • standard math Standard graph theory assumptions that nodes and weighted edges can represent fusing locations and their spatial-process relationships
    Invoked when defining the weighted network representation of the manufacturing process.
invented entities (2)
  • Weighted network representation of the manufacturing process no independent evidence
    purpose: To fuse geometric, process, and sensing data into a unified structure for modeling 3D interactions
    Newly introduced in the paper as the input structure for the transformer.
  • Dual-attention graph transformer no independent evidence
    purpose: To separately capture within-node feature dependencies and cross-node neighborhood interactions for quality learning
    New architecture proposed in the paper.

pith-pipeline@v0.9.1-grok · 5784 in / 1284 out tokens · 22374 ms · 2026-06-27T16:59:13.555665+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

32 extracted references · 21 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Arnab, A., Dehghani, M., Heigold, G., Sun, C., Luˇ ci´ c, M., Schmid, C.: ViVit: A Video Vision Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6836–6846 (2021)

    2021

  2. [2]

    Additive Manufacturing15, 1–11 (2017) https: //doi.org/10.1016/j.addma.2017.02.001

    Abdelrahman, M., Reutzel, E.W., Nassar, A.R., Starr, T.L.: Flaw detection in powder bed fusion using optical imaging. Additive Manufacturing15, 1–11 (2017) https: //doi.org/10.1016/j.addma.2017.02.001

    work page doi:10.1016/j.addma.2017.02.001 2017

  3. [3]

    IEEE Transactions on Automation Science and Engineering (2025) https://doi.org/10.1109/TASE.2025

    Alenezi, D.F., Shi, J., Li, J.: Graph-based variation propagation network for modeling and prediction of hybrid multi-stage manufacturing systems. IEEE Transactions on Automation Science and Engineering (2025) https://doi.org/10.1109/TASE.2025. 3560174

    work page doi:10.1109/tase.2025 2025

  4. [4]

    In: International Conference on Machine Learning, pp

    Chen, D., O’Bray, L., Borgwardt, K.: Structure-aware transformer for graph represen- tation learning. In: International Conference on Machine Learning, pp. 3469–3489 (2022). PMLR

    2022

  5. [5]

    Chaos: An Interdisciplinary Journal of Nonlinear Science28(8) (2018) https://doi.org/10.1063/1.5024917

    Chen, C.-B., Yang, H., Kumara, S.: Recurrence network modeling and analysis of spatial data. Chaos: An Interdisciplinary Journal of Nonlinear Science28(8) (2018) https://doi.org/10.1063/1.5024917

    work page doi:10.1063/1.5024917 2018

  6. [6]

    arXiv preprint arXiv:2012.09699 (2020)

    Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)

    arXiv 2012

  7. [7]

    arXiv preprint arXiv:2010.11929 (2020)

    Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  8. [8]

    Communications Materials1(1), 92 (2020) https://doi.org/10.1038/s43246-020-00094-y 20

    DePond, P.J., Fuller, J.C., Khairallah, S.A., Angus, J.R., Guss, G., Matthews, M.J., Martin, A.A.: Laser-metal interaction dynamics during additive manufactur- ing resolved by detection of thermally-induced electron emission. Communications Materials1(1), 92 (2020) https://doi.org/10.1038/s43246-020-00094-y 20

    work page doi:10.1038/s43246-020-00094-y 2020

  9. [9]

    Measurement Science and Technology28(4), 044005 (2017) https://doi.org/10.1088/1361-6501/aa5c4f

    Grasso, M., Colosimo, B.M.: Process defects and in situ monitoring methods in metal powder bed fusion: a review. Measurement Science and Technology28(4), 044005 (2017) https://doi.org/10.1088/1361-6501/aa5c4f

    work page doi:10.1088/1361-6501/aa5c4f 2017

  10. [10]

    Journal of Manufacturing Systems 80, 176–193 (2025) https://doi.org/10.1016/j.jmsy.2025.02.016

    Hussong, M., Ruediger-Flore, P., Klar, M., Kloft, M., Aurich, J.C.: Selection of manu- facturing processes using graph neural networks. Journal of Manufacturing Systems 80, 176–193 (2025) https://doi.org/10.1016/j.jmsy.2025.02.016

    work page doi:10.1016/j.jmsy.2025.02.016 2025

  11. [11]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016)

    2016

  12. [12]

    doi:10.1080/24725854.2017.1408165 , issn =

    Bian, L.: In-situ monitoring of melt pool images for porosity prediction in directed energy deposition processes. IISE Transactions51(5), 437–455 (2019) https://doi. org/10.1080/24725854.2017.1417656

    work page doi:10.1080/24725854.2017.1417656 2019

  13. [13]

    Quality and Relia- bility Engineering International33(8), 2003–2022 (2017) https://doi.org/10.1002/ qre.2163

    Kan, C., Yang, H.: Dynamic network monitoring and control of in situ image profiles from ultraprecision machining and biomanufacturing processes. Quality and Relia- bility Engineering International33(8), 2003–2022 (2017) https://doi.org/10.1002/ qre.2163

    2003

  14. [14]

    Journal of Intelligent Manufacturing33(2), 457–471 (2022) https://doi.org/10.1007/s10845-021-01842-8

    Larsen, S., Hooper, P.A.: Deep semi-supervised learning of dynamics for anomaly detection in laser powder bed fusion. Journal of Intelligent Manufacturing33(2), 457–471 (2022) https://doi.org/10.1007/s10845-021-01842-8

    work page doi:10.1007/s10845-021-01842-8 2022

  15. [15]

    Journal of Manufacturing Processes115, 240–255 (2024) https: //doi.org/10.1016/j.jmapro.2024.01.082

    Li, Q., Huang, T., Liu, J., Tan, L.: Time-series vision transformer based on cross space- time attention for fault diagnosis in fused deposition modelling with reconstruction of layer-wise data. Journal of Manufacturing Processes115, 240–255 (2024) https: //doi.org/10.1016/j.jmapro.2024.01.082

    work page doi:10.1016/j.jmapro.2024.01.082 2024

  16. [16]

    Journal of research of the National Institute of Standards and Technology125, 125027 (2020) https://doi.org/10.6028/ jres.125.027

    Lane, B., Yeung, H.: Process monitoring dataset from the additive manufacturing metrology testbed (ammt): Overhang part x4. Journal of research of the National Institute of Standards and Technology125, 125027 (2020) https://doi.org/10.6028/ jres.125.027

    2020

  17. [17]

    Journal of Manufacturing Science and Engineering145(11), 111008 (2023) https: //doi.org/10.1115/1.4063234

    Lee, H., Yang, H.: Digital twinning and optimization of manufacturing process flows. Journal of Manufacturing Science and Engineering145(11), 111008 (2023) https: //doi.org/10.1115/1.4063234

    work page doi:10.1115/1.4063234 2023

  18. [18]

    Additive Manufacturing48, 102449 (2021) https://doi.org/10.1016/j.addma.2021

    Mozaffar, M., Liao, S., Lin, H., Ehmann, K., Cao, J.: Geometry-agnostic data-driven thermal modeling of additive manufacturing processes using graph neural networks. Additive Manufacturing48, 102449 (2021) https://doi.org/10.1016/j.addma.2021. 102449

    work page doi:10.1016/j.addma.2021 2021

  19. [19]

    Additive Manufacturing Letters , author =

    Ogoke, F., Pak, P., Myers, A., Quirarte, G., Beuth, J., Malen, J., Farimani, A.B.: Deep learning for melt pool depth contour prediction from surface thermal images 21 via vision transformers. Additive Manufacturing Letters11, 100243 (2024) https: //doi.org/10.1016/j.addlet.2024.100243

    work page doi:10.1016/j.addlet.2024.100243 2024

  20. [20]

    In: 2025 IEEE Conference on Artificial Intelligence (CAI), pp

    Uhrich, B., Rahm, E.: MPGT: Multimodal physics-constrained graph transformer learning for hybrid digital twins. In: 2025 IEEE Conference on Artificial Intelligence (CAI), pp. 26–32 (2025). https://doi.org/10.1109/CAI64502.2025.00011 . IEEE Veliˇ ckovi´ c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv prep...

    work page doi:10.1109/cai64502.2025.00011 2025

  21. [21]

    Advances in neural information processing systems30(2017)

    Polosukhin, I.: Attention is all you need. Advances in neural information processing systems30(2017)

    2017

  22. [22]

    npj Advanced Manufacturing2(1), 15 (2025) https://doi.org/10.1038/s44334-025-00025-0

    Wang, X.Q., Jin, Z., Zheng, B., Gu, G.X.: Transformer-based approach for printing quality recognition in fused filament fabrication. npj Advanced Manufacturing2(1), 15 (2025) https://doi.org/10.1038/s44334-025-00025-0

    work page doi:10.1038/s44334-025-00025-0 2025

  23. [23]

    Journal of Indus- trial Information Integration45, 100795 (2025) https://doi.org/10.1016/j.jii.2025

    Wang, K., Lin, H., Fang, N., Xu, J., Zhang, S., Tan, J., Qin, J., Liang, X.: Cross- patch graph transformer enforced by contrastive information fusion for energy demand forecasting towards sustainable additive manufacturing. Journal of Indus- trial Information Integration45, 100795 (2025) https://doi.org/10.1016/j.jii.2025. 100795

    work page doi:10.1016/j.jii.2025 2025

  24. [24]

    Applied Network Science9(1), 25 (2024) https://doi.org/10.1007/ s41109-024-00637-z

    Yang, B., Gharebhaygloo, M., Rondi, H.R., Hortis, E., Lostalo, E.Z., Huang, X., Ercal, G.: Comparative analysis of course prerequisite networks for five midwestern pub- lic institutions. Applied Network Science9(1), 25 (2024) https://doi.org/10.1007/ s41109-024-00637-z

    2024

  25. [25]

    Journal of Manufacturing Science and Engineering140(3), 031014 (2018) https://doi.org/10.1115/1.4037891

    Yao, B., Imani, F., Sakpal, A.S., Reutzel, E.W., Yang, H.: Multifractal analysis of image profiles for the characterization and detection of defects in additive manufac- turing. Journal of Manufacturing Science and Engineering140(3), 031014 (2018) https://doi.org/10.1115/1.4037891

    work page doi:10.1115/1.4037891 2018

  26. [26]

    Yang, Z., Lu, Y., Yeung, H., Lane, B., Van Handel, N.: A fully registered in-situ and ex- situ dataset for metal powder bed fusion additive manufacturing: Data processing, feature extraction, registration, and uncertainties (2025) https://doi.org/10.6028/ NIST.AMS.100-69

    2025

  27. [27]

    In: 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE), pp

    Yang, Z., Lu, Y., Yeung, H., Krishnamurty, S.: Investigation of deep learning for real-time melt pool classification in additive manufacturing. In: 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE), pp. 640– 647 (2019). https://doi.org/10.1109/COASE.2019.8843291

    work page doi:10.1109/coase.2019.8843291 2019

  28. [28]

    Journal of Manufacturing Science and Engineering145(10), 101004 (2023) https://doi.org/10.1115/1.4057013 22

    Yang, H., Reijonen, J., Revuelta, A.: Multiresolution quality inspection of layerwise builds for metal 3d printer and scanner. Journal of Manufacturing Science and Engineering145(10), 101004 (2023) https://doi.org/10.1115/1.4057013 22

    work page doi:10.1115/1.4057013 2023

  29. [29]

    Proceedings of the IEEE109(4), 347–376 (2020) https://doi.org/10.1109/JPROC.2020.3034519

    Kumara, S.: Six-sigma quality management of additive manufacturing. Proceedings of the IEEE109(4), 347–376 (2020) https://doi.org/10.1109/JPROC.2020.3034519

    work page doi:10.1109/jproc.2020.3034519 2020

  30. [30]

    The International Journal of Advanced Manufacturing Technology136(9), 4055–4066 (2025) https://doi.org/10

    Zhang, S., Lu, Y., Yang, H.: Multiscale basis modeling of 3d melt-pool morphologi- cal variations for manufacturing process monitoring. The International Journal of Advanced Manufacturing Technology136(9), 4055–4066 (2025) https://doi.org/10. 1007/s00170-024-13377-2

    2025

  31. [31]

    Contactless video-based photoplethysmography tech- nique comparison investigating pulse transit time estimation of arterial blood pressure,

    Zhang, S., Yang, H.: Spatial modeling and analysis of human traffic and infectious virus spread in community networks. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 2286–2289 (2021). https://doi.org/10.1109/EMBC46164.2021.9630798

    work page doi:10.1109/embc46164.2021.9630798 2021

  32. [32]

    A New Approach to Linear Filtering and Prediction Problems

    Zhang, S., Yang, H., Yang, Z., Lu, Y.: Engineering-guided Deep Learning of Melt-pool Dynamics for Additive Manufacturing Quality Monitoring. Journal of Computing and Information Science in Engineering24(10) (2024) https://doi.org/10.1115/1. 4066026 23

    work page doi:10.1115/1 2024