build gnn surrogates

Traditional surrogate models struggle to simultaneously optimize both the structure and parameters of supply chains and lack generalization capabilities for graph-structured systems. This work presents the first systematic exploration of graph neural networks (GNNs) as surrogate models for supply chain design, introducing a programmatically generated supply chain graph dataset and leveraging a custom simulation library, SupplyNetPy, to produce large-scale training data for end-to-end differentiable modeling. The proposed approach enables performance prediction at both node and network levels, effectively balancing predictive accuracy and computational cost. By supporting gradient-based topology optimization and sensitivity analysis, this method establishes a foundation for novel design-space exploration strategies, significantly enhancing the efficiency of supply chain configuration and optimization.

Traditional graph neural networks (GNNs) struggle to disentangle the spatial influence of individual features in multi-variable forecasting on unstructured scientific meshes, severely limiting model interpretability. To address this, we propose FIGNN—a novel GNN architecture featuring (i) feature-specific pooling that isolates spatial dependencies per predicted variable, and (ii) an interpretable-error-aligned learnable spatial mask regularization that enforces physically consistent attribution of spatial importance. By incorporating physics-informed training objectives, FIGNN decouples feature representations while preserving predictive fidelity. Evaluated on the SPEEDY atmospheric model and BFS fluid dynamics benchmark, FIGNN achieves state-of-the-art prediction accuracy and uncovers physically plausible, feature-specific spatial patterns—e.g., distinct pressure- versus vorticity-driven flow structures. This advances scientific surrogate modeling by jointly improving both predictive performance and mechanistic interpretability.

Uncertainty quantification (UQ) in large-scale partial differential equation (PDE) simulations demands efficient, high-fidelity surrogate models—particularly for complex geometries and unstructured meshes. Method: We propose a physics-informed domain decomposition graph neural network (GNN) framework: the computational domain is partitioned into subdomains on unstructured grids; local GNN submodels are trained in parallel; and graph attention mechanisms coupled with transfer learning enable cross-subdomain knowledge sharing. Contribution/Results: Compared to global GNN surrogates, our approach significantly improves training efficiency and generalization capability. In high-resolution ice-sheet simulations, it achieves high-accuracy full-field ice velocity prediction while reducing training time by an order of magnitude. The framework provides a scalable, physics-consistent, and data-efficient surrogate modeling paradigm for large-scale scientific simulations on complex geometries.

Graph neural networks (GNNs) suffer from degraded training efficiency and generalization performance due to noise in graph topology and node features. To address this, we propose a model-sensitivity-based framework for identifying and actively selecting non-robust samples—marking the first work to quantify sample-level non-robustness as a principled criterion for constructing training subsets. Our method integrates gradient sensitivity analysis, localized perturbation assessment, and a Top-k sampling strategy, enabling seamless integration into standard GNN training pipelines (e.g., GCN, GAT) without architectural modifications. Evaluated on multiple benchmark graph datasets, our approach accelerates training by 23%–37%, improves average classification accuracy by 1.8%, and achieves significantly superior robustness compared to state-of-the-art baselines.

This work investigates the node-level generalization capability of graph neural networks (GNNs) for interpolating bandlimited functions defined on Euclidean cubes, focusing on exact recovery under label sparsity. Methodologically, it establishes the first theoretical connection between GNN architectures and the classical sampling theorem, integrating tools from graph signal processing, spectral graph theory, and bandlimited function analysis to construct a GNN with asymptotically optimal complexity. Theoretically, it proves that ε-accurate interpolation requires only O_d((log(1/ε))^d) parameters and labeled samples—exponentially fewer than the O_d((1/ε)^d) dependence typical of standard neural networks. This result is the first to reveal an intrinsic advantage of GNNs in modeling bandlimited signals over structured domains, providing a new theoretical foundation and design paradigm for efficient, principled GNN development.

Existing graph neural networks (GNNs) have been employed as surrogate models in computational fluid dynamics, yet their application to dynamic structural mechanics—particularly wave-dominated problems—remains unexplored. This work introduces GNSS, the first GNN-based surrogate for dynamic structural simulation. Methodologically, GNSS features: (1) node-fixed local coordinate systems to eliminate numerical cancellation in velocity estimation; (2) a sign-aware loss function to suppress phase drift during long-horizon rollout; and (3) wavelength-aware graph connectivity to optimize topological structure. Built upon an encode-process-decode architecture, it integrates finite-difference-based velocity estimation with physics-informed graph construction. Evaluated on a 50-kHz pulsed excitation beam benchmark, GNSS achieves hundreds of high-fidelity time steps with strong generalization to unseen dynamics. It outperforms explicit finite-element solvers significantly in inference speed while preserving spatiotemporal accuracy.

This work addresses the challenge of computationally expensive circuit simulations required for predicting voltage distributions in high-temperature superconducting magnets by proposing an efficient graph neural network–based surrogate model. The magnet’s equivalent circuit is represented as a graph, where a message-passing mechanism integrates circuit topology, material properties, and operating current information. Kirchhoff’s current law is incorporated as a physics-informed regularization constraint to enhance physical consistency. The resulting model exhibits topology-agnostic behavior, enabling zero-shot generalization and few-shot fine-tuning across diverse configurations. Evaluated over the design space, it achieves a mean absolute percentage error (MAPE) of 4.3% on average, allowing rapid inference of current redistribution and local operating conditions. This capability renders the model suitable for both design exploration and real-time monitoring of superconducting magnets.

This work addresses critical challenges in graph neural networks (GNNs) concerning representation learning, generalization capability, and adversarial robustness. The authors propose a novel GNN framework that enhances representation learning through an improved graph shift operator, boosts generalization via an effective graph data augmentation strategy, and strengthens adversarial robustness by integrating orthogonalization with a controlled noise injection mechanism. Extensive experiments demonstrate that the proposed method consistently outperforms existing approaches across diverse tasks and scenarios. The framework not only achieves superior overall performance but also offers new theoretical insights and practical pathways for deploying robust and generalizable GNN models.

This study addresses the automatic identification of synthetic graph generation models. We propose a hybrid classification framework integrating interpretable graph-theoretic features with graph neural networks (GNNs). On a large-scale heterogeneous graph dataset spanning five canonical generative models, we systematically co-optimize engineered topological features—both node- and graph-level—with six mainstream GNNs (GCN, GAT, GIN, GraphSAGE, GTN, etc.), revealing the critical role of message-passing mechanisms in discriminative performance. We incorporate ensemble random forest-based feature selection and Optuna-driven hyperparameter optimization, and establish the first benchmark evaluation suite for graph generation model classification. Experiments show GraphSAGE and GTN achieve 98.5% accuracy; t-SNE and UMAP visualizations demonstrate clear inter-class separation; GAT-based models underperform due to limitations in global structural modeling; and SVM baselines confirm the necessity of message passing for effective classification.

Finite element analysis (FEA) suffers from high computational cost, hindering its integration into iterative design optimization. Method: This paper proposes an efficient physics-informed surrogate model that integrates graph neural networks (GNNs) with physical constraints. We formulate a physics-informed neural network (PINN) framework grounded in the Navier–Cauchy equations and adopt a curriculum learning strategy for two-stage training—data-driven pretraining followed by physics-constrained fine-tuning. We systematically compare attention-enhanced GNN architectures, including GCN, MPNN, and Graph Transformer. Results: The Graph Transformer achieves a 2.6% relative L² error; the MPNN-PINN variant offers the best trade-off among accuracy, parameter count, and inference speed; and physics constraints reduce prediction error by 11.3% in high signal-to-noise ratio scenarios. This work provides the first systematic empirical validation of GNNs’ superiority in structural mechanics surrogate modeling, establishing a new paradigm for interpretable, generalizable FEA surrogates.