MeshODENet: Graph-Informed ODE Simulator

• MeshODENet is a graph-informed neural ODE model that integrates mesh topology with continuous-time dynamics to simulate physical systems.
• The architecture embeds the entire GNN as the derivative function of an ODE, ensuring stable long-term predictions even under large deformations.
• Empirical results show MeshODENet significantly reduces error accumulation compared to autoregressive surrogates like MeshGraphNet.

Searching arXiv for MeshODENet and closely related mesh/ODE papers to ground the article. MeshODENet is a graph-informed neural ordinary differential equation neural network for simulating mesh-based physical systems. It combines the spatial inductive bias of graph neural networks with the continuous-time formulation of Neural ODEs, and is designed for long-term, stable simulation of deformable mechanical systems discretized on meshes, especially under large, nonlinear deformation. In the formulation introduced under the title “MeshODENet: A Graph-Informed Neural Ordinary Differential Equation Neural Network for Simulating Mesh-Based Physical Systems,” the entire GNN is embedded as the derivative function of an ODE defined on an augmented mechanical state, while the graph itself has fixed connectivity and dynamic features derived from the evolving geometry and state (Liu et al., 22 Sep 2025).

1. Conceptual setting and scope

MeshODENet is situated in surrogate modeling for computational mechanics. Its target setting is the simulation of physical systems represented on a discretized mesh, where traditional numerical solvers such as FEM, FVM, DER, and DEP are robust but can be computationally prohibitive for many-query tasks. The motivating concern is not merely mesh awareness in the abstract, but long-horizon prediction for systems whose geometry and physical attributes evolve continuously in time (Liu et al., 22 Sep 2025).

The method is framed against three limitations of common learned surrogates. First, MLPs require fixed-size inputs and discard mesh topology when the mesh is flattened. Second, CNNs assume regular Euclidean grids and do not naturally accommodate irregular meshes. Third, autoregressive GNN surrogates such as MeshGraphNet are well aligned with local physical interactions but accumulate error over long rollouts because they consume their own previous predictions step by step. MeshODENet addresses this last limitation by learning a continuous-time vector field rather than a one-step transition operator (Liu et al., 22 Sep 2025).

This design is especially relevant for mesh-based mechanics because prior graph ODE applications often focus on graphs whose topology and attributes are mostly static, whereas the regime considered here has fixed connectivity but state-dependent node and edge attributes. A plausible implication is that MeshODENet should be understood less as a generic graph ODE and more as a continuous-time simulator whose graph features must be reconstructed from the current physical configuration at each vector-field evaluation.

2. Mathematical formulation and architecture

At time $t$, the physical system is represented as a graph

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$

with node set $\mathcal{V}$, edge set $\mathcal{E}$, node-feature matrix $\mathbf{H}_v$, and edge-feature matrix $\mathbf{H}_e$. Physical interactions are assumed to be undirected, but each undirected edge is implemented as two directed edges so that information can flow both ways during message passing (Liu et al., 22 Sep 2025).

The GNN has an encoder–processor–decoder structure. Raw node and edge features are first embedded by separate MLPs,

$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$

The processor then applies $L$ message-passing layers. Edge messages are computed as

$$\mathbf{m}_{e_{ij}}^{(l+1)} = \phi_e^{(l)}\!\left( \tilde{\mathbf{h}}_{e_{ij}}^{(l)}, \tilde{\mathbf{h}}_{v_i}^{(l)}, \tilde{\mathbf{h}}_{v_j}^{(l)} \right),$$

and node states are updated by aggregating incoming messages with a permutation-invariant operator $\bigoplus$: $$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$0 A decoder MLP then outputs nodal acceleration,

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$1

Collectively,

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$2

In the reported experiments, MeshODENet uses only one message-passing layer, whereas the MeshGraphNet baseline uses 15 layers in the 1-D case and 30 layers in the 2-D case (Liu et al., 22 Sep 2025).

The neural ODE is defined on the augmented mechanical state

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$3

The state derivative is

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$4

so the GNN acts as the continuously re-evaluated acceleration field of a second-order mechanical system. Rollout is given by

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$5

or, equivalently,

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$6

Although adaptive solvers such as Dormand–Prince are discussed as possibilities, the implementation uses explicit fourth-order Runge–Kutta (RK4), which evaluates the GNN four times per integration step. Gradients are computed by backpropagating through the ODE solver via the adjoint sensitivity method (Liu et al., 22 Sep 2025).

3. State variables, graph features, and physical encoding

A distinctive feature of MeshODENet is that its graph features are explicitly physics-informed. For the 1-D elastic beam, the node feature at node $$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$7 is

$$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$8

These encode nodal velocity, Young’s modulus, local geometric information through the angle $$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{H}_v, \mathbf{H}_e),$$9, and external force. The angular information is represented by $\mathcal{V}$0 and $\mathcal{V}$1 to avoid angle wrapping discontinuities. The corresponding edge feature is

$\mathcal{V}$2

where $\mathcal{V}$3 and $\mathcal{V}$4 are undeformed edge length and direction, while $\mathcal{V}$5 and $\mathcal{V}$6 are the current length and direction. This makes deformation explicit through the contrast between reference and current geometry (Liu et al., 22 Sep 2025).

For the 2-D cantilever plate, the angle feature is removed and a node-type feature is introduced to distinguish fixed boundary nodes from free nodes. In both cases, the graph connectivity remains fixed, but the node and edge attributes are dynamic because they are recomputed from the current state (Liu et al., 22 Sep 2025).

The training data are generated from high-fidelity mechanics solvers rather than from the neural model itself. For the rod, the semi-discrete equation of motion is

$\mathcal{V}$7

with elastic energy

$\mathcal{V}$8

stretching energy

$\mathcal{V}$9

and bending energy

$\mathcal{E}$0

These equations motivate the surrogate problem and define the source of reference trajectories, but MeshODENet does not directly impose them as a physics residual during training (Liu et al., 22 Sep 2025).

4. Training procedure and benchmark problems

Training is performed on full rollouts from an initial state rather than on one-step targets. The loss is imposed on integrated positions over the predicted trajectory: $\mathcal{E}$1 Only this trajectory-level position MSE is explicitly reported; no auxiliary velocity loss, acceleration loss, or physics residual loss is added. The formulation is intended to provide a more stable training signal and to encourage long-term physical consistency (Liu et al., 22 Sep 2025).

Two benchmark problems are reported.

Problem	Ground-truth simulator	Data generation
1-D elastic rod falling in fluid	DER	71 training trajectories, 14 test trajectories; $\mathcal{E}$2 for training and $\mathcal{E}$3 for testing; duration 20 s; sampling interval 0.1 s
2-D cantilever plate with large deformation	DEP	30 training trajectories, 6 test trajectories; $\mathcal{E}$4; duration 1.5 s; 500 steps; step size 0.003 s

The 1-D rod has length $\mathcal{E}$5 m, $\mathcal{E}$6 vertices, rod radius $\mathcal{E}$7 mm, viscosity $\mathcal{E}$8 Pa·s, and external forces consisting of gravity minus buoyancy plus viscous drag. The 2-D plate has length $\mathcal{E}$9 m, width $\mathbf{H}_v$0 m, thickness $\mathbf{H}_v$1 m, density $\mathbf{H}_v$2 kg/m$\mathbf{H}_v$3, one fixed end, gravity loading, and mild velocity-proportional damping (Liu et al., 22 Sep 2025).

Implementation uses PyTorch and PyTorch Geometric with Adam. For the 1-D rod, the hidden dimension is 128, MeshODENet uses one message-passing layer, training runs for 400 epochs, the initial learning rate is $\mathbf{H}_v$4, the learning rate decays by a factor of 10 every 100 epochs, and weight decay is $\mathbf{H}_v$5. For the 2-D plate, the hidden dimension is kept consistent, MeshODENet again uses one message-passing layer, training runs for 600 epochs, the initial learning rate is $\mathbf{H}_v$6, and the learning rate decays by a factor of 10 at epochs 200 and 400 (Liu et al., 22 Sep 2025).

5. Reported empirical behavior, accuracy, and limitations

The principal empirical claim is improved long-term stability relative to the autoregressive MeshGraphNet baseline. In the 1-D rod benchmark, RMSE averaged over 14 test trajectories is reported at selected rollout steps as follows (Liu et al., 22 Sep 2025):

Step	MeshODENet	MGN
1	$\mathbf{H}_v$7	$\mathbf{H}_v$8
30	$\mathbf{H}_v$9	$\mathbf{H}_e$0
60	$\mathbf{H}_e$1	$\mathbf{H}_e$2
90	$\mathbf{H}_e$3	$\mathbf{H}_e$4
120	$\mathbf{H}_e$5	$\mathbf{H}_e$6
150	$\mathbf{H}_e$7	$\mathbf{H}_e$8
180	$\mathbf{H}_e$9	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$0

The reported relative increase of MGN over MeshODENet exceeds 9000% by step 180. In the 2-D plate benchmark, RMSE averaged over 6 test trajectories is (Liu et al., 22 Sep 2025):

Step	MeshODENet	MGN
1	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$1	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$2
25	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$3	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$4
50	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$5	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$6
150	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$7	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$8
250	$$\tilde{\mathbf{h}}_{v_i}^{(0)} = \mathrm{MLP}_{v,\mathrm{enc}}(\mathbf{h}_{v_i}), \qquad \tilde{\mathbf{h}}_{e_{ij}}^{(0)} = \mathrm{MLP}_{e,\mathrm{enc}}(\mathbf{h}_{e_{ij}}).$$9	$L$0
350	$L$1	$L$2

Here the reported relative error increase of MGN reaches over 12,800% by step 350. The qualitative interpretation given for both tasks is that MGN can remain reasonable over short horizons but eventually suffers severe error accumulation, whereas MeshODENet maintains coherent deformation and structural integrity over the rollout horizon (Liu et al., 22 Sep 2025).

The reported limitations are equally important. The main stated limitation is substantial memory consumption associated with graph data for large meshes. The framework also depends on careful feature engineering, particularly as physical complexity increases. High-dimensional systems with varied external loads may create data sparsity relative to the complexity of the force–response map. A specific failure mode is delayed response to sudden, violent dynamic changes: in the most flexible rod case, the baseline MGN can be momentarily more accurate at the very beginning of a sharply changing transient. The experiments also assume fixed graph connectivity and do not address topological change, fracture, remeshing, or contact-induced connectivity updates. Benchmark diversity remains limited to 1-D rods in fluid and 2-D plates under gravity, and the text gives little detail on sensitivity to solver choice or step size in the backward pass (Liu et al., 22 Sep 2025).

The abstract claims substantial computational speed-ups over traditional solvers, but the body does not provide a timing table or explicit speed-up factor. This suggests that the efficiency claim is motivated by the avoidance of repeated nonlinear solves rather than by a fully quantified runtime benchmark in the reported experiments.

6. Relation to adjacent methods and nomenclature

The name MeshODENet is easily conflated with several nearby lines of work, but the distinctions are technically important. MeshODENet is a continuous-time graph simulator for deformable meshes, in which a GNN evaluated on the current graph state is embedded directly as the derivative function of a Neural ODE (Liu et al., 22 Sep 2025).

It is distinct from MeshODE, which is a pairwise CAD deformation framework built around an ODE-based bijective deformation map optimized per source–target pair without prespecified correspondences. MeshODE addresses non-rigid CAD registration and shape interpolation, not rollout prediction of dynamical mesh-based physical systems (Huang et al., 2020). It is also distinct from MeshONet, whose “O” denotes operator learning rather than ordinary differential equations; MeshONet is a dual-branch, shared-trunk operator-learning architecture for structured mesh generation and is not a neural ODE model (Xiao et al., 21 Jan 2025).

A closer conceptual relative is DDOT, “Diffeomorphic mesh Deformation via an efficient Optimal Transport metric,” which uses a neural ODE to deform cortical surface meshes and replaces Chamfer-based supervision with sliced Wasserstein distance over probabilistic mesh representations, especially oriented varifolds. DDOT is therefore a neural ODE on explicit mesh vertices, but its task is cortical surface reconstruction rather than simulation of physical time evolution (Le et al., 2023). By contrast, MAgNET is a graph U-Net surrogate for nonlinear finite-element simulations that maps mesh-based loads to equilibrium displacements in one shot; it is mesh-aware but not continuous-time and not ODE-driven (Deshpande et al., 2022). HodgeNet occupies yet another neighboring space: it learns discrete differential operators on triangle meshes and uses their spectral behavior to build descriptors, but it does not define continuous-depth latent dynamics or solve a neural ODE on mesh states (Smirnov et al., 2021).

These comparisons clarify the technical identity of MeshODENet. It is neither a mesh-generation operator net nor a static graph surrogate nor a pairwise deformation optimizer. It is specifically a graph-informed neural ordinary differential equation model for long-horizon simulation on mesh-based physical systems, with fixed connectivity, dynamic graph features, and trajectory-level supervision on positions (Liu et al., 22 Sep 2025).