On the Potential of Graph Neural Networks as Metamodels for Supply Chain Optimization: Dataset, Architectures, and Directions

Neha Karanjkar; Tushar Lone

REVIEW 3 major objections 4 minor 24 references

On the Potential of Graph Neural Networks as Metamodels for Supply Chain Optimization: Dataset, Architectures, and Directions

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper tries to establish that a single graph neural network can act as a fast, differentiable metamodel for supply-chain simulation, predicting performance on unseen networks more than twenty times larger than those seen in training.

desk verdict Useful first step on GNN metamodels for supply chains, with a strong dataset contribution, but the headline generalization claim overreaches: only three of five metrics hold at 20x scale. read the letter →

arxiv 2607.16769 v1 pith:PPH33DMX submitted 2026-07-18 cs.LG

Tushar Lone , Neha Karanjkar This is my paper

classification cs.LG

keywords graphneuralnetworkssimulationmetamodelsupplychainoptimizationdiscrete-eventmessagepassinggeneralizationacrosstopologiessyntheticdatasetsimulation-based

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 authors ask whether GNNs can replace slow discrete-event simulation inside supply-chain optimization loops, where standard metamodels (Kriging, polynomial, MLP) fail because they ignore network topology. They build a custom message-passing GNN that ingests node parameters, edge attributes, and graph structure, and train it on a new public dataset of 100,000 programmatically generated two-echelon supply chains. The model predicts five performance metrics (carry cost, spend cost, transport cost, revenue, profit) with R² ≈ 0.99 on held-out networks and keeps high accuracy on networks more than twenty times the training size, though transport and carry costs degrade at the largest shifted scales. The payoff of the claim is that a differentiable surrogate opens gradient-based joint optimization of both parameters and topology, which black-box simulation cannot support.

What carries the argument

The engine is a custom message-passing layer (a variant of the MPNN framework) that updates each node's hidden state by summing messages built from the concatenation of sender and receiver features and the edge feature vector, passed through a ReLU MLP. Three design choices do the heavy lifting: vector-valued edge features (transport cost and lead time) enter the message, back edges with a direction flag let retailers influence upstream distributors, and a learned node-type embedding specializes behavior by role. Sum aggregation over the immediate neighborhood keeps the prediction semantics local and size-independent, which the paper identifies as the main reason the model transfers to much

What would settle it

Train the same architecture on the released dataset, then evaluate it on a held-out set of stochastic-demand or multi-echelon/cyclic networks generated with the same simulator (or on any real supply-chain data with ground-truth performance). If R² for profit or revenue drops substantially below 0.99—or if transport-cost R² turns negative at moderate sizes—the claimed cross-size generalization is not robust beyond the paper's restricted class. Concretely, the transport-cost collapse beyond roughly 350 nodes under shifted parameters is already a partial falsifier that the paper acknowledges.

Watch

Extended reading notes

Core claim

The paper's central claim is that a structure-aware GNN—an edge-conditioned, node-type-aware message-passing architecture with back edges and sum aggregation—learns a smooth mapping from supply-chain graph (topology, node parameters, edge costs/lead times) to steady-state performance, and that this mapping transfers across graph sizes. Evaluated on held-out and larger unseen networks, the single trained model attains R² ≈ 0.99 for all five node-level metrics on test and moderately larger networks; profit, revenue, and spend cost remain near 0.99 on networks up to 421 nodes (more than 20× the largest training network), while inventory carry cost degrades to about 0.8–0.4 and transport cost be

Load-bearing premise

The entire result rests on the assumption that SupplyNetPy's synthetic two-echelon, deterministic-demand, min-max (s,S) tree networks are representative enough of real supply chains that a surrogate trained on them will predict performance on real, stochastic, possibly cyclic supply networks; the paper explicitly calls this a deliberate simplification and fixes the policy while excluding stochasticity, node failures, and perishability.

Editorial extensions

If this is right

A single trained GNN can replace simulation inside an optimization loop: one forward pass takes about 1 ms versus roughly 6–13 seconds per simulation run, a speed-up of about three orders of magnitude.
Because the surrogate is differentiable end-to-end and continuous in graph structure as well as parameters, it can in principle support gradient-based joint optimization of topology and parameters, continuous relaxation of edge existence, and sensitivity analysis.
Network-level performance follows from pooling node-level predictions with no extra computation, so the same model serves both node and system queries.
Generalization across sizes means a model trained on small graphs can evaluate candidate designs far outside the training distribution (up to 20× larger), reducing dataset generation cost.
Accuracy-compute trade-offs are tunable: training-set size and message-passing depth govern the R²/cost frontier, with accuracy saturating near 60,000 training points.

Reading between the lines

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

The one-hop locality and sum aggregation suggest the approach should extend to deeper multi-echelon chains only if message-passing depth grows with the longest path; otherwise effects beyond immediate neighbors will be missed — a testable extension of the paper's own argument.
The marked degradation of transport and carry cost on shifted/large networks hints that edge-cost prediction may need global or multi-hop context (e.g., path-length or bottleneck features), not just local messages; this is an inference, not a paper claim.
Because the training data is limited to tree topologies with deterministic demand and a fixed (s,S) policy, the strongest reading of the result is restricted to that class; a quick test would be to generate cyclic or stochastic networks with the same simulator and check whether R²≈0.99 survives.
The paper's own future direction—continuous relaxation of edge probabilities for topology optimization—could be probed immediately by checking whether the trained model's gradients with respect to edge features align with brute-force edge-removal simulation, a validation the paper does not perform.

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.

Desk Editor's Note

read the letter

Short version: this is a solid, honest initial study with a genuinely useful dataset, but the abstract and C1 claim more than the evidence supports. The stress-test note is right. Figure 2 shows transport-cost R2 goes negative beyond roughly 350 nodes and carry cost drops to ~0.8 at 421 nodes under training parameter ranges, so 'more than twenty times larger' is only true for profit, revenue, and spend cost. The body text does acknowledge a validity boundary in Q3, but the contribution list and abstract don't carry that qualification. That's an overstatement, not a fatal flaw—the core idea survives.

What's genuinely new: a 100k-graph synthetic dataset of two-echelon supply chains, open and programmatically generated, and a custom edge-conditioned MPNN that reads node and edge features with back edges to propagate downstream influence. The ablation study is clean and informative: back edges and edge features are load-bearing, node-type embeddings less so. They also report tail errors (P95) and compare against GCN/GraphSAGE, showing the custom design beats standard layers by a wide margin. The cross-size generalization, even if only on three metrics, is still noteworthy—R2 near 0.99 on networks much larger than training for the strongest metrics.

Soft spots, in proportion: the generalization overclaim is the biggest issue and needs a careful rewrite. Beyond that, the empirical evidence would be stronger with multiple seeds and error bars—they train once, and run-to-run variation is mentioned but not quantified. There's no classical metamodel baseline (e.g., GPR or MLP on flattened features) on these same graphs; they cite prior work to motivate that such baselines fail, but a direct comparison would sharpen the claim. The deterministic demand and fixed (s,S) policy are explicit simplifications, and the authors are honest about them—that's fine for a foundation paper, but it limits immediate practical relevance.

The paper is coherent on its own terms and the authors engage with limitations honestly. It deserves peer review; the dataset alone justifies referee time. With revisions that qualify the generalization claim and add error bars, it would be a solid contribution for the supply-chain simulation community. I'd bring it to a reading group focused on GNN surrogates, and I'd cite it for the dataset if working in this area.

Referee Report

3 major / 4 minor

Summary. The paper proposes GNNs as differentiable metamodels for two-echelon supply chain simulation. It contributes a custom edge-conditioned, node-type-aware message-passing GNN trained on a new synthetic dataset of 100,000 SupplyNetPy-generated network configurations, predicting five node-level metrics (carry cost, spend cost, transport cost, revenue, profit). The authors report R²≈0.99 on held-out and moderately larger networks, analyze architecture choices via ablations, compare the trained model against GCN/GraphSAGE baselines, measure accuracy/compute trade-offs, and identify a validity boundary for transport and carry costs on much larger or parameter-shifted networks. The paper explicitly frames Q5 (differentiable optimization) as future work, not a demonstrated result.

Significance. If the empirical claims are accepted with the appropriate qualifications, this is a useful foundation: it is among the first public datasets for graph-structured supply chain metamodeling, and the ablations give concrete guidance about which architectural components matter (edge features, back edges, node-type embeddings). The explicit quantification of a validity boundary is a strength. However, the headlined cross-size generalization result is overstated as written for two of the five metrics, and the absence of repeated-seed error bars weakens the quantitative comparisons. The deterministic, simplified simulation setting limits external validity, though the paper is transparent about this. The contribution is best read as an initial demonstration on a controlled class of supply chains, with the optimization agenda still open.

major comments (3)

[§4.4, Fig. 2; C1 and Abstract] The unqualified claim that the model 'demonstrates generalization ... to networks more than twenty times larger than those seen during training' is contradicted by the paper's own Figure 2. The largest training network has at most 16 nodes, while the scaling set reaches 421 nodes, i.e., more than 20×. Under training parameter ranges, transport-cost R² turns negative beyond roughly 350 nodes and carry cost drops to about 0.8 at 421 nodes; under shifted ranges carry degrades to about 0.4 and transport to about −2.3. Thus only profit, revenue, and spend cost support the 20× claim. C1 and the abstract must carry the precise caveat: the generalization holds for 3 of 5 metrics and only through a stated size/parameter validity boundary, or the experiments must be strengthened to match the claim.
[§4.3–4.4, Table 2] All reported R², MAE, P95, and architecture-variant comparisons come from a single trained model, a single data split, and a single initialization. Table 2 is a one-run comparison. With differences such as 0.989 vs 0.987, the claim that only the full model and the no-node-type-embedding variant keep every metric above 0.98 is not robust to seed and split variation. The manuscript should report means and standard deviations over several seeds/splits, at least for the main results and the architecture ablations, before quantitative claims such as 'only the full model keeps all metrics above 0.98' are made.
[§3.2, §4.1] The high R² is partly a consequence of deterministic, noise-free targets, as the paper itself notes in §4.4. The simulation space is restricted to two-echelon tree networks, deterministic demand, a fixed min-max policy, and single-run labels, with stochasticity, disruptions, and perishability excluded. Since the stated goal is supply-chain metamodeling, this is a load-bearing external-validity gap. The revision should either narrow the title/claims to this simplified class (e.g., 'deterministic two-echelon supply chains') or include a stochastic-label experiment (targets as means over replications) to show that the demonstrated accuracy survives label noise.

minor comments (4)

[§3.2] The term 'steady-state performance metrics' is used but the simulation run length, warm-up period, and convergence criterion are not specified. Please provide these details so readers can reproduce the ground-truth labels.
[Abstract / C2] The dataset is described as 'publicly released' but no URL or DOI is given in the text; only the SupplyNetPy GitHub repository is cited. Add a direct dataset link or accession identifier.
[§5.1] When sweeping dataset size D and MLP depth L, the manuscript does not state which evaluation set is used to compute R² (test split vs unseen larger networks). Please specify, since the operational interpretation of Figure 3 depends on it.
[§4.2] The phrase 'A single GNN layer suffices' could be misread as conflicting with the five-layer MLP used inside the message-passing layer. Clarify that this refers to the number of message-passing hops, not the depth of the MSG MLP.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation chain; the GNN is trained on and tested against an independent (author-built but external) simulator, and self-citations are motivational/tooling, not load-bearing.

full rationale

The paper's derivation chain is: SupplyNetPy defines an input-to-output simulation mapping; 100,000 supply-chain graphs are generated; a GNN is trained with MSE to predict node-level metrics from node and edge features; and accuracy is measured on an 80/20 split of that dataset plus separately generated larger and parameter-shifted networks. The labels come from the simulator, which is not a function of the GNN, so the held-out and cross-size R2 results are genuine out-of-sample predictions rather than fitted parameters renamed as predictions. The only author-linked inputs are the SupplyNetPy ground-truth library ([13]) and a prior report that fixed-topology metamodels failed to generalize ([14]); neither is the conclusion of this paper. The architecture claims are independently supported by the within-paper comparisons against GCN/GraphSAGE and by ablations (removing back edges, edge features, and node-type embeddings). Therefore no circular step is present. The overbroad C1 scaling claim is a correctness/internal-consistency issue rather than a circularity issue: the paper's own Figure 2 and text state that transport-cost R2 turns negative beyond about 350 nodes and carry cost degrades to roughly 0.8 at 421 nodes, so the 'more than twenty times larger' generalization claim is supported only for profit, revenue, and spend cost. That should be addressed as a reporting/validity-boundary concern, not as circular reasoning.

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

The central empirical claim depends on the validity of the authors' simulator as ground truth, the local-neighborhood assumption that justifies a one-layer model, and the representativeness of the hand-chosen synthetic generation ranges. These are domain assumptions, not standard mathematical facts, and are stated openly in the paper.

assumptions (3)

domain assumption SupplyNetPy's discrete-event simulation produces valid ground-truth performance metrics for supply chains.
Used throughout Section 3; the simulator is not validated against real-world or independent simulation data.
domain assumption A node's steady-state performance depends primarily on its immediate neighbors, so a single message-passing layer is sufficient.
Section 4.3: 'A single GNN layer suffices, since each node's performance depends primarily on its immediate neighbors.' This is what enables cross-size generalization.
ad hoc to paper Uniform random generation of tree-structured two-echelon networks yields a representative training distribution for the metamodel.
Section 3.2 defines the generation process; the ranges in Table 1 are chosen by the authors and are not derived from real data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On the Potential of Graph Neural Networks as Metamodels for Supply Chain Optimization: Dataset, Architectures, and Directions." pith.science (2026). https://pith.science/paper/PPH33DMX

@misc{pith2026260716769,
  author       = {Pith},
  title        = {Pith review of: On the Potential of Graph Neural Networks as Metamodels for Supply Chain Optimization: Dataset, Architectures, and Directions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPH33DMX}},
  note         = {Machine review of arXiv:2607.16769}
}

read the original abstract

Graph Neural Networks (GNNs) have emerged as a powerful, differentiable class of learning models for graph-structured systems. Their ability to generalize across topologies opens the prospect of a surrogate for combined structural and parametric optimization, which classical metamodels cannot offer. Supply chains are a natural target, yet the use of GNN surrogates for supply chain problems is largely unexplored. This paper lays the foundation, presents initial steps, and discusses key research directions. As a foundation, we formulate the problem and create a large public training dataset of programmatically generated supply chain graphs with input parameters and steady-state performance metrics obtained using our SupplyNetPy simulation library. As initial steps, we explore GNN architectures that work well as surrogates for node- and network-level predictions, and analyze their accuracy-compute trade-off against simulation. Most importantly, we outline the exciting directions this opens, namely gradient-based optimization over topology, fast design-space exploration, and sensitivity analysis.

Figures

Figures reproduced from arXiv: 2607.16769 by the authors.

**Figure 2.** R 2 of profit, transport cost, and inventory carry cost versus network size, under (left) training parameter ranges and (right) shifted parameter ranges. Dashed and dotted lines mark R 2 = 1 and 0. or node-type embedding, exceed a mean predictor but stay well short of a usable surrogate, trailing the custom architecture by a wide margin. The single-factor rows pinpoint the decisive components, namely the edge-condit… view at source ↗

**Figure 3.** Accuracy (R 2 ) versus computational cost (per-evaluation time) for the GNN metamodel and the simulation model. and Q4 are addressed by the empirical results, Q2 by the dataset contribution, and Q5 defines the research agenda that this paper sets out and that we are pursuing. Q1: can GNNs work well as surrogates for SC performance prediction, and what architectures are suitable? We presented a custom message-passin… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 5 linked inside Pith

[1]

GNN- based probabilistic supply and inventory predictions in supply chain networks.arXiv preprint arXiv:2404.07523, 2024

Hyung-il Ahn, Young Chol Song, Santiago Olivar, Hershel Mehta, and Naveen Tewari. GNN- based probabilistic supply and inventory predictions in supply chain networks.arXiv preprint arXiv:2404.07523, 2024

arXiv 2024
[2]

Causal dynamic Bayesian networks for simulation metamodeling

Pracheta Amaranath, Peter J Haas, David Jensen, and Sam Witty. Causal dynamic Bayesian networks for simulation metamodeling. In2023 Winter Simulation Conference (WSC), pages 746–757. IEEE, 2023

2023
[3]

Maria Chiara Angelini and Federico Ricci-Tersenghi. Modern graph neural networks do worse than classical greedy algorithms in solving combinatorial optimization problems like maximum indepen- dent set.Nature Machine Intelligence, 5(1):29–31, 2023

2023
[4]

Enhanced simulation metamodeling via graph and generative neural networks

Wang Cen and Peter J Haas. Enhanced simulation metamodeling via graph and generative neural networks. In2022 Winter Simulation Conference (WSC), pages 2748–2759. IEEE, 2022

2022
[5]

Efficient hybrid simulation optimization via graph neural network meta- modeling

Wang Cen and Peter J Haas. Efficient hybrid simulation optimization via graph neural network meta- modeling. In2023 Winter Simulation Conference (WSC), pages 3541–3552. IEEE, 2023

2023
[6]

Linlin Feng. Joint optimization algorithm for vehicle scheduling and supply chain inventory manage- ment based on multi-agent deep reinforcement learning.Neural Computing and Applications, pages 1–27, 2025

2025
[7]

Neural message passing for quantum chemistry

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. InInternational Conference on Machine Learning, pages 1263–1272. PMLR, 2017

2017
[8]

Research on optimization and management of supply chain networks based on graph neural networks

Chuanyou Guo. Research on optimization and management of supply chain networks based on graph neural networks. In2024 6th International Academic Exchange Conference on Science and Technology Innovation (IAECST), pages 233–236. IEEE, 2024

2024

Show all 24 references

[9]

Inductive representation learning on large graphs

Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in Neural Information Processing Systems, 30, 2017. 12

2017
[10]

Erd ˝os goes neural: An unsupervised learning framework for combinatorial optimization on graphs.Advances in Neural Information Processing Systems, 33:6659– 6672, 2020

Nikolaos Karalias and Andreas Loukas. Erd ˝os goes neural: An unsupervised learning framework for combinatorial optimization on graphs.Advances in Neural Information Processing Systems, 33:6659– 6672, 2020

2020
[11]

Semi-supervised classification with graph convolutional networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[12]

DIFFIM: Differentiable influ- ence minimization with surrogate modeling and continuous relaxation

Junghun Lee, Hyunju Kim, Fanchen Bu, Jihoon Ko, and Kijung Shin. DIFFIM: Differentiable influ- ence minimization with surrogate modeling and continuous relaxation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18128–18136, 2025

2025
[13]

SupplyNetPy Github repository.https://github.com/ SupplyChainSimulation/SupplyNetPy, 2024

Tushar Lone and Neha Karanjkar. SupplyNetPy Github repository.https://github.com/ SupplyChainSimulation/SupplyNetPy, 2024

2024
[14]

An open tool-set for simulation, design-space explo- ration and optimization of supply chains and inventory problems

Tushar Lone, P Lekshmi, and Neha Karanjkar. An open tool-set for simulation, design-space explo- ration and optimization of supply chains and inventory problems. InSIMULTECH, pages 432–439, 2023

2023
[15]

Development of an open-source library for supply chain modeling and opti- mization

Tushar Mohanrao Lone. Development of an open-source library for supply chain modeling and opti- mization. InProceedings of the 38th ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, pages 73–74, 2024

2024
[16]

Metamodel-based quantile estimation for hedging control of manufacturing systems

Giulia Pedrielli and Russell R Barton. Metamodel-based quantile estimation for hedging control of manufacturing systems. In2019 Winter Simulation Conference (WSC), pages 452–463. IEEE, 2019

2019
[17]

Combinatorial optimization with physics-inspired graph neural networks.Nature Machine Intelligence, 4(4):367–377, 2022

Martin JA Schuetz, J Kyle Brubaker, and Helmut G Katzgraber. Combinatorial optimization with physics-inspired graph neural networks.Nature Machine Intelligence, 4(4):367–377, 2022

2022
[18]

SupplyGraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

Azmine Toushik Wasi, MD Islam, and Adipto Raihan Akib. SupplyGraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

2024 arXiv
[19]

Semi-supervised graph convolutional neural network based classification for auto parts inventory management

Fangpeng Weng, Bing Guo, Xinhua Suo, Xin Wu, Yuchuan Hu, and Qin Zheng. Semi-supervised graph convolutional neural network based classification for auto parts inventory management. In2022 IEEE 24th Int Conf on High Performance Computing & Communications; 8th Int Conf on Data S...

2022
[20]

Metamodel-assisted sensitivity analysis for controlling the impact of input uncertainty

Wei Xie, Bo Wang, and Pu Zhang. Metamodel-assisted sensitivity analysis for controlling the impact of input uncertainty. In2019 Winter Simulation Conference (WSC), pages 3681–3692. IEEE, 2019

2019
[21]

DAG-GNN: DAG structure learning with graph neural networks

Yue Yu, Jie Chen, Tian Gao, and Mo Yu. DAG-GNN: DAG structure learning with graph neural networks. InInternational Conference on Machine Learning, pages 7154–7163. PMLR, 2019

2019
[22]

An analytics-driven approach to enhancing supply chain visibility with graph neural networks and federated learning.arXiv preprint arXiv:2503.07231, 2025

Ge Zheng and Alexandra Brintrup. An analytics-driven approach to enhancing supply chain visibility with graph neural networks and federated learning.arXiv preprint arXiv:2503.07231, 2025

2025 arXiv
[23]

A machine learning approach for enhancing supply chain visibility with graph-based learning.Supply Chain Analytics, page 100135, 2025

Ge Zheng and Alexandra Brintrup. A machine learning approach for enhancing supply chain visibility with graph-based learning.Supply Chain Analytics, page 100135, 2025

2025
[24]

Iterative multi- agent reinforcement learning: A novel approach toward real-world multi-echelon inventory optimiza- tion.arXiv preprint arXiv:2503.18201, 2025

Georg Ziegner, Michael Choi, Hung Mac Chan Le, Sahil Sakhuja, and Arash Sarmadi. Iterative multi- agent reinforcement learning: A novel approach toward real-world multi-echelon inventory optimiza- tion.arXiv preprint arXiv:2503.18201, 2025. 13

2025 arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

GNN- based probabilistic supply and inventory predictions in supply chain networks.arXiv preprint arXiv:2404.07523, 2024

Hyung-il Ahn, Young Chol Song, Santiago Olivar, Hershel Mehta, and Naveen Tewari. GNN- based probabilistic supply and inventory predictions in supply chain networks.arXiv preprint arXiv:2404.07523, 2024

arXiv 2024

[2] [2]

Causal dynamic Bayesian networks for simulation metamodeling

Pracheta Amaranath, Peter J Haas, David Jensen, and Sam Witty. Causal dynamic Bayesian networks for simulation metamodeling. In2023 Winter Simulation Conference (WSC), pages 746–757. IEEE, 2023

2023

[3] [3]

Maria Chiara Angelini and Federico Ricci-Tersenghi. Modern graph neural networks do worse than classical greedy algorithms in solving combinatorial optimization problems like maximum indepen- dent set.Nature Machine Intelligence, 5(1):29–31, 2023

2023

[4] [4]

Enhanced simulation metamodeling via graph and generative neural networks

Wang Cen and Peter J Haas. Enhanced simulation metamodeling via graph and generative neural networks. In2022 Winter Simulation Conference (WSC), pages 2748–2759. IEEE, 2022

2022

[5] [5]

Efficient hybrid simulation optimization via graph neural network meta- modeling

Wang Cen and Peter J Haas. Efficient hybrid simulation optimization via graph neural network meta- modeling. In2023 Winter Simulation Conference (WSC), pages 3541–3552. IEEE, 2023

2023

[6] [6]

Linlin Feng. Joint optimization algorithm for vehicle scheduling and supply chain inventory manage- ment based on multi-agent deep reinforcement learning.Neural Computing and Applications, pages 1–27, 2025

2025

[7] [7]

Neural message passing for quantum chemistry

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. InInternational Conference on Machine Learning, pages 1263–1272. PMLR, 2017

2017

[8] [8]

Research on optimization and management of supply chain networks based on graph neural networks

Chuanyou Guo. Research on optimization and management of supply chain networks based on graph neural networks. In2024 6th International Academic Exchange Conference on Science and Technology Innovation (IAECST), pages 233–236. IEEE, 2024

2024