M4GN: Mesh-Based Hierarchical Graph Network

• M4GN is a mesh-based multi-segment hierarchical graph network that efficiently simulates PDEs while maintaining fine-scale accuracy and handling long-range interactions.
• Its three-tier architecture integrates a micro-level GNN for local dynamics, intermediate physics-aware segmentation, and a macro-level transformer for global reasoning.
• The method overcomes over-smoothing challenges and reduces computational cost, achieving significant RMSE improvements and faster inference compared to baselines.

M4GN most commonly denotes “Mesh-based Multi-segment Hierarchical Graph Network,” a hierarchical graph-neural surrogate for PDE simulations on unstructured meshes. It was introduced to address the high cost, long propagation paths, and over-smoothing or oversquashing associated with deep mesh-based GNNs, while preserving fine-scale accuracy and respecting mesh topology, geometry, and physical discontinuities. The method is organized as a three-tier, segment-centric hierarchy: a micro-level GNN operating on the original mesh, an intermediate segmentation stage that builds contiguous and dynamically consistent segments, and a macro-level transformer that performs efficient global reasoning over segment tokens (Lei et al., 12 Sep 2025).

1. Definition and problem setting

M4GN is designed for mesh-based graph neural network surrogates for PDE simulations. In this setting, the computational domain is discretized as a mesh graph whose nodes are mesh vertices and whose edges encode mesh connectivity; in Lagrangian cases, additional world edges can encode proximity-based contact or collision. The central difficulty addressed by M4GN is that standard mesh GNNs require deep message passing to propagate information over large meshes or long-range physical couplings, which increases computational cost and can degrade representations through over-smoothing and oversquashing (Lei et al., 12 Sep 2025).

The method targets two obstacles that remain even in hierarchical GNNs. The first is coarse graph construction: a useful hierarchy must respect topology, geometry, and physical discontinuities such as holes, cracks, and interfaces, instead of merging nodes merely because they are close in Euclidean space. The second is retaining fine-scale accuracy: aggressive pooling can irreversibly remove high-frequency information, and sequence-based segment aggregators such as GRUs can introduce order sensitivity, vanishing gradients, and additional computational cost (Lei et al., 12 Sep 2025).

Within this formulation, M4GN preserves the original mesh at the micro level, forms physically coherent segments at an intermediate level, and reasons across those segments with a transformer at the macro level. This suggests a deliberate compromise between two desiderata that are usually in tension in learned PDE surrogates: local fidelity and long-range efficiency.

2. Three-tier architecture and information flow

The architecture consists of four functional blocks: a micro-level GNN, an intermediate segmentation stage, a macro-level transformer, and a node-level decoder. The micro-level module operates directly on the mesh graph and captures fine-scale local dynamics using limited message-passing depth. The segmentation stage partitions the mesh into $K$ contiguous segments and computes segment adjacency and positional information. The macro-level transformer treats those segments as tokens and exchanges long-range information across them with $O(K^2)$ attention cost, where $K \ll N$. The decoder then maps node features to next-step predictions with an $L_2$ supervision objective (Lei et al., 12 Sep 2025).

A concise view of the hierarchy is as follows:

Tier	Representation	Role
Micro	Mesh nodes and edges	Capture local dynamics
Intermediate	Contiguous segments	Compress physically coherent regions
Macro	Segment tokens	Reason over long-range couplings

At the micro level, initial node and edge encodings are defined by

$$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$

For $l=1,\dots,L$, message passing updates node and edge states as

$$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$

$$\mathbf{h}_{ij,t}^{M,l} = f_{e}^{l}(\mathbf{h}_{ij,t}^{M,l-1}, \mathbf{h}_{i,t}^{l-1}, \mathbf{h}_{j,t}^{l-1}), \quad \mathbf{h}_{ij,t}^{W,l} = f_{e}^{l}(\mathbf{h}_{ij,t}^{W,l-1}, \mathbf{h}_{i,t}^{l-1}, \mathbf{h}_{j,t}^{l-1}).$$

This micro-level encoder–process stage is explicitly intended to keep message-passing depth moderate, with long-range propagation delegated to the macro layer (Lei et al., 12 Sep 2025).

At the macro level, one embedding $\mathbf{h}_{S_i}$ is formed for each segment and passed through multi-head self-attention blocks with pre-LayerNorm. After $L_S$ transformer blocks, the segment feature is dispatched back to nodes by concatenation,

$O(K^2)$0

which creates the “information bridge” between local and global dynamics. A plausible implication is that M4GN shortens effective information paths without replacing the original mesh graph by a fully pooled surrogate (Lei et al., 12 Sep 2025).

3. Hybrid mesh segmentation

A defining feature of M4GN is its hybrid segmentation strategy, which combines a graph partitioner with a SLIC-style refinement procedure. The segmentation policy is written as $O(K^2)$1, producing $O(K^2)$2 segments $O(K^2)$3 with disjoint node sets; overlap is optionally enabled through a ring-expansion parameter $O(K^2)$4, with $O(K^2)$5 in the overlap experiments (Lei et al., 12 Sep 2025).

The first stage uses METIS, which produces $O(K^2)$6 non-overlapping, balanced, contiguous segments by multilevel coarsening, partitioning, uncoarsening, and refinement. The second stage performs SLIC-style refinement guided by physics, using node descriptors of the form

$O(K^2)$7

where $O(K^2)$8 are coordinates, $O(K^2)$9 are obstacle or rigid-body cues, and $K \ll N$0 are modal-decomposition features. The segment centroid is

$K \ll N$1

Assignment is based on the distance

$K \ll N$2

where $K \ll N$3 controls compactness (Lei et al., 12 Sep 2025).

The modal features differ by physical regime. For solids, M4GN uses structural modes from

$K \ll N$4

and for fluids it uses Laplacian eigenfunctions from

$K \ll N$5

Per-node modal descriptors are formed by retaining the first $K \ll N$6 modes. Obstacle-distance transforms

$K \ll N$7

are also considered in feature variants such as SLIC-OD, SLIC-OD$K \ll N$8, SLIC-MD, and SLIC-MDOD$K \ll N$9; the reported best overall segmentation variant is SLIC-MDOD$L_2$0, which combines modal features with exponentially weighted obstacle distance (Lei et al., 12 Sep 2025).

This segmentation mechanism is central to M4GN’s claim of being segment-centric rather than merely multiscale. The segments are intended to be contiguous, geometry-faithful, and dynamically consistent, so that the transformer operates on physically meaningful coarse tokens rather than arbitrary clusters.

4. Segment encoding, macro reasoning, and training objective

M4GN uses a permutation-invariant segment encoder rather than a recurrent segment aggregator. Segment encoding is given by

$L_2$1

where $L_2$2 are the node features after $L_2$3 micro-level message-passing steps and $L_2$4 is an MLP. The paper emphasizes that this Deep Sets-style average pooling is order-insensitive and has complexity $L_2$5 per segment, compared with $L_2$6 and long backpropagation chains for GRU-based alternatives (Lei et al., 12 Sep 2025).

Optional positional encodings can be added. At the segment level, the adjacency matrix is defined as

$L_2$7

and random-walk structural encodings can be injected as

$L_2$8

Node-level positional encodings may also be incorporated into the input node states (Lei et al., 12 Sep 2025).

The macro transformer applies multi-head self-attention over the $L_2$9 segment tokens. Its attention cost is $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$0, which is substantially lighter than node-level transformers because $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$1. The paper notes that M4GN uses a fully connected segment graph by default, though topology-aware masks based on $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$2 can be introduced when needed (Lei et al., 12 Sep 2025).

Training is supervised with per-node next-step prediction. The objective is

$$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$3

The model is then rolled out autoregressively to produce trajectories, with evaluation on 1-step, 50-step, and full-trajectory metrics. The training setup explicitly does not use curriculum learning or teacher forcing; instead it relies on direct next-step training plus rollout evaluation. Regularization includes Gaussian input-noise perturbations and a correction loss following Godwin et al. (Lei et al., 12 Sep 2025).

5. Computational characteristics and implementation

The computational decomposition of M4GN is explicit. The micro-level encoder–process stage has learnable complexity

$$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$4

while the macro transformer contributes

$$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$5

Preprocessing consists mainly of modal decomposition and hybrid segmentation. The former is described as sparse eigen-solves that scale approximately $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$6 for fixed $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$7, and the latter combines METIS multilevel partitioning with local SLIC refinement, whose iteration scales as $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$8 (Lei et al., 12 Sep 2025).

The reported implementation uses a micro GNN with $$\mathbf{h}_{i,t}^{0} = f_{n}(\mathbf{x}_{i,t}), \quad \mathbf{h}_{ij,t}^{M,0} = f_{e}^M(\mathbf{e}_{ij,t}^M), \quad \mathbf{h}_{ij,t}^{W,0} = f_{e}^W(\mathbf{e}_{ij,t}^W).$$9 message-passing steps, MLPs with 3 layers and hidden size 128, ReLU activations, LayerNorm at the end of encoders, and a 3-layer decoder MLP without LayerNorm. The macro transformer uses 4 layers, 8 heads, and pre-LayerNorm. Training employs an initial learning rate of $l=1,\dots,L$0 decaying to $l=1,\dots,L$1 over the final half of training, with 2M steps for CylinderFlow and DeformingPlate and 1M steps for DeformingBeam. The experiments use batch size 8, P100 GPUs, and distributed PyTorch (Lei et al., 12 Sep 2025).

Dataset-specific segmentation choices are also fixed. CylinderFlow uses $l=1,\dots,L$2 and $l=1,\dots,L$3, with overlap and positional encoding enabled and 6 modes for modal features. DeformingPlate uses $l=1,\dots,L$4 and $l=1,\dots,L$5, no overlap, no positional encoding, and only the first structural mode. DeformingBeam uses $l=1,\dots,L$6 and $l=1,\dots,L$7, overlap on, positional encoding off, and the first mode only. World-edge radii are 0.01 for DeformingPlate and 0.002 for DeformingBeam (Lei et al., 12 Sep 2025).

The method is explicitly intended for large meshes and/or long-range interactions, especially in settings where flat GNNs over-smooth or become slow, and in cases that require strong mesh fidelity together with efficient global coupling. This suggests a natural use case in learned surrogate modeling for solid mechanics and contact-rich Lagrangian dynamics.

6. Benchmarks, metrics, and empirical performance

M4GN is evaluated on several benchmark datasets: CylinderFlow, DeformingPlate, DeformingBeam, and DeformingBeam-Large, with supplementary results on EAGLE and FlagSimple. Baselines include GCN, g-U-Net, MeshGraphNets (MGN), BSMS-GNN, and EAGLE (Lei et al., 12 Sep 2025).

The primary numerical metric is RMSE for next-step and rollout prediction. For Lagrangian settings, the paper also reports mesh-quality metrics including Hausdorff $l=1,\dots,L$8, Chamfer $l=1,\dots,L$9, Mesh Continuity (MC), and Aspect Ratio error (AR)). Segmentation quality is analyzed with **Conductance, Edge Cut Ratio, and Silhouette score, which are used to relate segment cohesion and separation to downstream predictive performance (Lei et al., 12 Sep 2025).

Selected reported outcomes are as follows:

Dataset	Reported gain
CylinderFlow	RMSE-all reduced by 36% vs second-best
DeformingPlate	RMSE-all reduced by 32% vs next-best
DeformingBeam	RMSE-all reduced by up to 56% vs baselines
Overall inference	Up to 22% faster than state-of-the-art baselines

The paper gives more specific comparisons. On CylinderFlow, M4GN achieves RMSE-all $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$0 versus $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$1 for the second-best model, corresponding to a 36% reduction. On DeformingPlate, it lowers RMSE-all by 32% compared to the next-best baseline and improves mesh quality, including a 36% reduction in $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$2 beyond EAGLE. On DeformingBeam, it reduces RMSE-all by up to 56% relative to baselines, and its $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$3 and $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$4 are reported as lower by approximately 50–70% versus EAGLE (Lei et al., 12 Sep 2025).

The paper further reports that M4GN’s error increases slowly with graph diameter, whereas baseline errors escalate rapidly, and that on DeformingBeam-Large it shows the smallest generalization gap, with RMSE-all reduced by 46% versus EAGLE and MC improved by 45%. Replacing EAGLE’s segmentation with M4GN’s segmentation improves EAGLE’s performance by 15–23% lower prediction error and 28–35% better mesh quality, indicating that the segmentation design is itself a substantial contributor, separate from the full hierarchical model (Lei et al., 12 Sep 2025).

A further empirical point concerns efficiency relative to numerical solvers. Compared to ground-truth solvers, the reported per-step speedups are $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$5–$$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$6, using COMSOL or OpenFOAM references. This does not imply exact solver replacement, but it situates M4GN in the established surrogate-modeling regime where rollout efficiency is part of the objective.

7. Ablations, limitations, and name ambiguity

The ablation studies isolate three design choices. First, physics-aware segmentation refinement matters: the SLIC-MDOD$$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$7 variant yields the best accuracy and mesh metrics, and modal plus obstacle-distance features consistently outperform geometry-only clustering. Second, the segment aggregator matters: average pooling with an MLP outperforms GRU aggregators in both accuracy and efficiency. Third, depth selection matters: a transformer with 4 self-attention layers and 8 heads is reported as robust, while 3–7 micro-level message-passing steps balance reach and over-smoothing; deeper micro propagation exhibits oversmoothing (Lei et al., 12 Sep 2025).

The limitations are also explicit. M4GN does not impose hard constraints on contact boundaries or guarantee physical consistency across segment interfaces; instead it relies on learned dynamics and segmentation fidelity. Segmentation hyperparameters such as $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$8, $$\mathbf{h}_{i,t}^{l} = f_{n}^{l}\Big(\mathbf{h}_{i,t}^{l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{M,l-1}, \sum_{j\in \mathrm{Adj}(i)}\mathbf{h}_{ij,t}^{W,l-1}\Big),$$9, overlap, and positional encoding require empirical tuning, with some sensitivity when extrapolating to substantially larger meshes. For fluids, modal decomposition based on Laplacian eigenmodes may fail to capture highly turbulent structures as effectively as snapshot-based modes, so the benefit of segmentation can be dataset-dependent (Lei et al., 12 Sep 2025).

The string “M4GN” is, however, not unique across current technical literature. In the available arXiv record, the exact canonical name M4GN refers to the mesh-surrogate model described above (Lei et al., 12 Sep 2025). By contrast, several papers use closely related but distinct names:

• MAGNET, the nearby-galaxy survey “Mechanisms Affecting Galaxies Nearby and Environmental Trends,” notes that “M4GN” is not defined in the paper and would only amount to a stylized shorthand for part of the program name if used informally (Vulcani et al., 16 Feb 2026).
• MAGNet, the multi-agent reinforcement-learning method “Multi-agent Graph Network,” does not use “M4GN” as an official alias; if the string appears elsewhere, it is described as an informal shorthand rather than the canonical name (Malysheva et al., 2020).
• The multimodal recommendation model MAGNET similarly uses the acronym MAGNET and not M4GN as its official title, although the detailed notes describe “MAGNET (M4GN)” as a naming variant in explanation (Dai et al., 24 Feb 2026).
• In meshless magnetohydrodynamics, the paper on modified-gradient methods states that the published technique is called MG, while “M4GN” may be encountered as a practitioner alias emphasizing exact $$\mathbf{h}_{ij,t}^{M,l} = f_{e}^{l}(\mathbf{h}_{ij,t}^{M,l-1}, \mathbf{h}_{i,t}^{l-1}, \mathbf{h}_{j,t}^{l-1}), \quad \mathbf{h}_{ij,t}^{W,l} = f_{e}^{l}(\mathbf{h}_{ij,t}^{W,l-1}, \mathbf{h}_{i,t}^{l-1}, \mathbf{h}_{j,t}^{l-1}).$$0 and use of the M4 kernel (Tu et al., 4 Mar 2026).
• In high-energy astrophysics, MAGN denotes misaligned AGN, but that usage is conceptually unrelated to M4GN as a graph-network architecture (Angioni et al., 2017).

Accordingly, in contemporary technical usage, M4GN is best treated as an overloaded label whose most explicit canonical occurrence is the Mesh-based Multi-segment Hierarchical Graph Network for dynamic simulations. In contexts involving nearby-galaxy surveys, multi-agent RL, multimodal recommendation, or meshless MHD, the term generally indicates an informal or non-canonical shorthand rather than the primary published name.