XNODE in PDEs, Systems & Explainable GNNs

• XNODE is a multi-domain term with distinct applications, including neural ODE‐based PDE solvers, heterogeneous runtime scheduling, and self-explaining GNN frameworks.
• In PDE solving, XNODE leverages explicit spatial dependency in its neural ODE structure to accelerate convergence and effectively incorporate initial and boundary data.
• In systems and graph learning, XNODE denotes both a fine-grained runtime abstraction for multi-XPU scheduling and a self-explaining node mechanism for GNNs.

Searching arXiv for XNODE-related papers to ground the article in the cited literature. I’m unable to directly access the arXiv search tool in this interface, so I will rely strictly on the arXiv records and detailed data provided. The term XNODE is used in multiple, technically distinct ways in the literature. In the PDE literature, it denotes a neural-ODE-based trial architecture for solving high-dimensional parabolic and stationary PDEs within weak adversarial network formulations (Oliva et al., 2021, Bertoluzza et al., 2023). In autonomous systems, it denotes a stage-level, multi-XPU programming abstraction for heterogeneous scheduling in latency-sensitive pipelines (Han et al., 13 Aug 2025). A separate line of work uses the closely related form X-Node for a self-explaining graph neural network framework in which each node generates an explanation as part of prediction (Sengupta et al., 14 Aug 2025). The shared label therefore does not identify a single unified method; rather, it names several architectures or abstractions introduced in different domains.

1. Terminological scope and disambiguation

In the PDE setting, XNODE is introduced as a model for the primal network in a weak adversarial network solver for high-dimensional parabolic PDEs (Oliva et al., 2021). The motivation is that standard DNN solvers “do not take into account the fundamentally different roles played by time and spatial variables in the solution,” whereas XNODE is “built on the neural ODE (NODE) model with additional spatial dependency” (Oliva et al., 2021). In later theoretical work, a related pseudo-time XNODE neural network class is analyzed for static PDE problems and incorporated into WAN discretizations (Bertoluzza et al., 2023).

In autonomous systems, XNODE is defined differently: it is a “fine-grained, multi-XPU abstraction whose granularity aligns exactly with stage boundaries of an algorithm” (Han et al., 13 Aug 2025). Here the term refers not to a neural architecture but to a runtime abstraction used by XAUTO for assignment and scheduling across CPUs, GPUs, DLAs, and related accelerators (Han et al., 13 Aug 2025).

In explainable GNNs, the term appears as X-Node, a “self-explaining GNN framework in which each node generates its own explanation as part of the prediction process” (Sengupta et al., 14 Aug 2025). This usage is conceptually separate from both the PDE and systems meanings.

A plausible implication is that any technical discussion of “XNODE” must specify the application domain before the term becomes meaningful.

2. XNODE in high-dimensional PDE solving

The PDE-oriented XNODE begins from the observation that, for fixed spatial coordinate $x$, the solution $t \mapsto u(t,x)$ may be viewed through an ODE in time (Oliva et al., 2021). The construction introduces a latent state $h(t)$ governed by a neural vector field and a linear readout:

$$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$

The parameter set is collected as $\Theta=(\theta_1,\theta_2,\tilde\theta)$, and for a time grid $\Pi_T=\{t_i\}_{i=0}^{n_T}$ the output is

$$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$

These definitions are given explicitly in the end-to-end description accompanying the parabolic PDE paper (Oliva et al., 2021).

A key distinguishing feature is the treatment of spatial dependence. Unlike “standard Residual-Nets or DNN solvers which treat $(t,x)$ symmetrically,” XNODE feeds $x$ explicitly into the vector-field network at every time step, so that the PDE’s spatial heterogeneity $a_{ij}(t,x)$, $t \mapsto u(t,x)$0, and $t \mapsto u(t,x)$1 “can be learned on the fly” (Oliva et al., 2021). The architecture is therefore organized into three parts: initialization, ODE block, and readout (Oliva et al., 2021).

The same basic idea is extended in the theoretical WAN paper to stationary PDEs by introducing an artificial or “pseudo-” time variable (Bertoluzza et al., 2023). There, one chooses a spatial coordinate, for example $t \mapsto u(t,x)$2, as pseudo-time,

$t \mapsto u(t,x)$3

and evolves

$t \mapsto u(t,x)$4

with the solution read out as $t \mapsto u(t,x)$5 (Bertoluzza et al., 2023). In practice this continuous dynamics is typically discretized by explicit Euler on a uniform grid, producing a short residual block repeated $t \mapsto u(t,x)$6 times (Bertoluzza et al., 2023).

3. Integration into weak adversarial networks

In the parabolic formulation, XNODE is used as the primal network inside a weak adversarial network framework (Oliva et al., 2021). On a possibly time-varying domain $t \mapsto u(t,x)$7, the weak solution satisfies

$t \mapsto u(t,x)$8

with the bilinear form and linear functional specified by the PDE coefficients and source term (Oliva et al., 2021). WAN reformulates the problem as

$t \mapsto u(t,x)$9

The parameterization used in XNODE-WAN is

• primal network: $h(t)$0,
• adversarial test-function network: $h(t)$1 (Oliva et al., 2021).

The loss components are

$h(t)$2

$h(t)$3

and the overall objective is the saddle-point problem

$h(t)$4

The optimization alternates gradient descent in the primal parameters and gradient ascent in the adversarial parameters after Monte Carlo approximation of the relevant integrals (Oliva et al., 2021).

The theoretical analysis in the later WAN paper places XNODE within a broader abstract discretization framework (Bertoluzza et al., 2023). That paper studies existence and stability of discrete solutions, proves quasi-best approximation estimates similar to Céa’s lemma, and proposes “stabilized WAN” and “shifted sWAN” formulations that avoid direct normalization (Bertoluzza et al., 2023). For coercive problems, Lemma 4.2 yields

$h(t)$5

while a generalized bound is given for non-symmetric problems under a discrete inf-sup condition (Bertoluzza et al., 2023). The paper states that the key requirement for the best approximation outcome is that the test network space satisfy the inf-sup condition and that “the method’s accuracy, however, is only determined by the space of the trial network” (Bertoluzza et al., 2023).

4. Initial and boundary data, including time-varying domains

For time-independent domains, the initial condition $h(t)$6 is built into XNODE via the initializer network, while boundary data $h(t)$7 enters through a penalty term in the WAN loss (Oliva et al., 2021). This separation is one of the stated reasons for preferring the architecture: it “naturally incorporates the initial condition in the ODE-initializer network” (Oliva et al., 2021).

For time-varying domains, the treatment is more involved. Each constant-$h(t)$8 path is split into subpaths between entry and exit times $h(t)$9 on $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$0 (Oliva et al., 2021). On each subpath, the ODE is re-initialized either from the initial datum $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$1 when $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$2 or from the boundary value $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$3 otherwise. The formulation given in the paper is

$$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$4

This re-initialization mechanism is specific to the parabolic XNODE-WAN construction (Oliva et al., 2021).

The theoretical WAN paper addresses essential boundary conditions through a different but related mechanism, denoted “CutWAN” (Bertoluzza et al., 2023). There the adversarial test functions are multiplied by a smooth cutoff $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$5 vanishing on $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$6, and a boundary penalty is added so that $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$7 exactly (Bertoluzza et al., 2023). The paper states in Lemma 5.2 that the resulting formulation admits the same quasi-best-approximation error bound (Bertoluzza et al., 2023).

5. Empirical behavior in PDE applications

The numerical experiments in the parabolic paper solve

$$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$8

in dimensions $$\frac{d}{dt} h(t) = \mathcal{N}^{\mathrm{vec}}_{\theta_2}(h(t), t, x), \qquad h(0)=\mathcal{N}^{\mathrm{init}}_{\theta_1}(h(x)) \in \mathbb{R}^{h}, \qquad \hat u(t,x)=\mathcal{L}_{\tilde\theta}(h(t)) \in \mathbb{R}.$$9, on hypercubes, balls, and a one-dimensional “hourglass” time-varying domain (Oliva et al., 2021). The reported metrics are relative $\Theta=(\theta_1,\theta_2,\tilde\theta)$0 error, time per epoch, number of epochs to reach error $\Theta=(\theta_1,\theta_2,\tilde\theta)$1, and total wall-clock time $\Theta=(\theta_1,\theta_2,\tilde\theta)$2 (Oliva et al., 2021).

A representative result is given for the 5D cube with $\Theta=(\theta_1,\theta_2,\tilde\theta)$3 (Oliva et al., 2021):

Method	Relative error	Training time profile
WAN (DNN-primal)	$\Theta=(\theta_1,\theta_2,\tilde\theta)$4	$\Theta=(\theta_1,\theta_2,\tilde\theta)$5 s/epoch, $\Theta=(\theta_1,\theta_2,\tilde\theta)$6 epochs, $\Theta=(\theta_1,\theta_2,\tilde\theta)$7 s
XNODE-WAN	$\Theta=(\theta_1,\theta_2,\tilde\theta)$8	$\Theta=(\theta_1,\theta_2,\tilde\theta)$9 s/epoch, $\Pi_T=\{t_i\}_{i=0}^{n_T}$0 epochs, $\Pi_T=\{t_i\}_{i=0}^{n_T}$1 s

The paper also states that in high-dimensional tests with $\Pi_T=\{t_i\}_{i=0}^{n_T}$2 and $\Pi_T=\{t_i\}_{i=0}^{n_T}$3, XNODE-WAN converges in $\Pi_T=\{t_i\}_{i=0}^{n_T}$4 while WAN requires $\Pi_T=\{t_i\}_{i=0}^{n_T}$5, and that similar gains hold on time-varying domains (Oliva et al., 2021). The accompanying summary characterizes the effect as “orders-of-magnitude speed-up and faster convergence in $\Pi_T=\{t_i\}_{i=0}^{n_T}$6 up to 64” (Oliva et al., 2021).

The theoretical WAN paper reports analogous empirical comparisons for static and parabolic problems, stating that a classical DNN trained by the same losses typically needs $\Pi_T=\{t_i\}_{i=0}^{n_T}$7 gradient-descent steps to reach a few-percent accuracy and may oscillate, whereas XNODE “routinely reaches $\Pi_T=\{t_i\}_{i=0}^{n_T}$8-training-error in a few $\Pi_T=\{t_i\}_{i=0}^{n_T}$9 steps, with much smaller oscillations” (Bertoluzza et al., 2023). This suggests that the improvement is not tied only to one PDE class but to the trial-space parameterization itself.

6. Architectural rationale and theoretical claims

The stated rationale for XNODE in PDEs is architectural rather than purely empirical. The model “enforces the ODE structure in $$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$0, so that the network only needs to learn first-order time-dynamics (no high-order derivatives),” naturally incorporates the initial condition, and supports adjoint-method backpropagation with “constant memory cost in time” (Oliva et al., 2021). The parabolic paper also claims “provable universal approximation” and cites Theorem 3 in its appendix (Oliva et al., 2021).

In the later WAN analysis, XNODE is described as a “continuous neural-ODE in an artificial time variable” and as a short residual network under explicit Euler discretization (Bertoluzza et al., 2023). The network parameterization used in the experiments consists of a small fully-connected initializer, a deeper fully-connected vector-field network with 8–12 layers, and a trainable affine readout (Bertoluzza et al., 2023). The authors emphasize that, within the WAN framework, accuracy is governed by the trial-network space provided the test-network space is sufficiently rich in the inf-sup sense (Bertoluzza et al., 2023).

A plausible implication is that the benefits attributed to XNODE arise from an inductive bias tailored to PDE evolution structure, rather than from increased parameter count alone. The data support this only indirectly, since the comparison is described as using “approximately the same number of free parameters” under the same loss formulations (Bertoluzza et al., 2023).

7. Other meanings of XNODE in the literature

Outside PDEs, XNODE denotes unrelated constructs. In autonomous applications, XNODE is a programming abstraction rather than a neural network (Han et al., 13 Aug 2025). It identifies a single algorithmic stage within a module’s data flow, and each XNODE may include multiple implementations on different XPUs, with one selected at runtime (Han et al., 13 Aug 2025). Through XAUTO, XNODEs are assigned to processors and scheduled by solving an ILP with constraints for assignment, precedence, WCET, and demand-bound feasibility, with the objective of minimizing maximal finish time (Han et al., 13 Aug 2025). On a driving perception pipeline, the paper reports 99th-percentile end-to-end latency of $$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$1 ms for ROS2 versus $$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$2 ms for XAUTO, a $$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$3 improvement (Han et al., 13 Aug 2025). This is a systems abstraction whose goals, equations, and evaluation criteria are entirely distinct from the PDE usage.

In graph learning, X-Node denotes a self-explaining GNN framework (Sengupta et al., 14 Aug 2025). Each node builds a structured context vector from seven cues—degree, clustering coefficient, 2-hop label agreement, eigenvector centrality, betweenness centrality, average edge weight, and community membership—and a two-layer MLP Reasoner maps this context to an explanation vector $$XNODE_\Theta(x,\Pi_T) = (\hat u(t_i,x))_{i=0,\dots,n_T}.$$4 (Sengupta et al., 14 Aug 2025). The explanation vector is used for classification, embedding reconstruction, and explanation alignment, and can be fed back into the model through a “text-injection” mechanism (Sengupta et al., 14 Aug 2025). Reported experiments use GCN, GAT, and GIN backbones on graph datasets derived from MedMNIST and MorphoMNIST (Sengupta et al., 14 Aug 2025). Although the name overlaps, this framework is neither derived from neural ODEs nor related to heterogeneous XPU scheduling.

The coexistence of these meanings is a potential source of confusion. In current arXiv usage, XNODE is therefore best understood as a domain-specific label adopted independently in PDE solvers and in runtime scheduling systems, while X-Node identifies a separate explainable-GNN method (Oliva et al., 2021, Han et al., 13 Aug 2025, Sengupta et al., 14 Aug 2025).