NeuroPDE: Neural Methods for PDEs

• NeuroPDE is a research field fusing neural networks with partial differential equations to model dynamic, spatiotemporal phenomena with methods like CNN dynamics and operator learning.
• It employs diverse strategies such as continuous-time formulations, mesh-based finite elements, residual minimization, and neuro-symbolic frameworks to enforce numerical structure.
• Advanced NeuroPDE approaches integrate traditional numerical methods, hybrid mechanistic learning, and stochastic or path-dependent extensions to enhance interpretability and performance.

Searching arXiv for recent and foundational papers on NeuroPDE / NeuralPDE and related neural PDE methods. NeuroPDE denotes a heterogeneous research area at the intersection of neural networks and partial differential equations. In the literature considered here, the term covers continuous-time neural models for PDE-governed dynamics, mesh-based and variational finite-neuron formulations, discretization-aware residual methods, operator-learning systems for deterministic and stochastic PDEs, path-dependent and non-Markovian extensions, neuro-symbolic frameworks, hardware realizations, and application-specific hybrids in neuroscience and scientific computing (Dulny et al., 2021, Khara et al., 2021, Mistani et al., 2022, Fu et al., 7 Jul 2025). A plausible implication is that “NeuroPDE” is best treated as a field label rather than a single canonical algorithm.

1. Scope and nomenclature

Within this literature, closely related labels are used in distinct ways. "NeuralPDE: Modelling Dynamical Systems from Data" formulates a continuous-time model for spatiotemporal fields by observing that the Method of Lines can be represented using convolutions, so convolutional neural networks become a natural parameterization of PDE dynamics (Dulny et al., 2021). "NeuroPDE: A Neuromorphic PDE Solver Based on Spintronic and Ferroelectric Devices" instead denotes a hardware architecture for Monte Carlo random-walk PDE solving that exploits intrinsic device stochasticity (Fu et al., 7 Jul 2025). Other papers use neither title but clearly belong to the same methodological cluster, including neural finite-element solvers, neural operator surrogates, path-dependent PDE solvers, and hybrid mechanistic-learning systems (Khara et al., 2021, Mundinger et al., 2024, Fang et al., 2023).

A recurring distinction across this body of work is between learning a single PDE instance and learning a family-level solution operator. The operator-learning viewpoint is explicit in Neural Parameter Regression, which defines

$$G:\mathcal{X}\supset K \to \mathcal{Y},\qquad u_0 \mapsto \big((t,x)\mapsto u(t,x)\big),$$

and in Neural SPDE, which learns solution operators for PDEs with possibly stochastic forcing (Mundinger et al., 2024, Salvi et al., 2021). By contrast, discretization-aware methods such as the Neural Bootstrapping Method remain solver-like and are trained on the instance rather than amortized across many PDEs (Mistani et al., 2022).

A second distinction concerns what is learned. Some methods learn the solution field directly; some learn a discrete operator or local stencil; some learn latent parameter dynamics; some learn constitutive or gating laws inside an otherwise classical PDE solver; and some search for closed-form symbolic expressions rather than numerical fields (Wu et al., 2023, Gaby et al., 2023, Gurung et al., 2024, Huang et al., 12 Feb 2026). This heterogeneity is central to the subject.

2. Core architectural families

Several architectural families recur across the literature.

Family	Representative method	Distinctive idea
Continuous-time CNN dynamics	NeuralPDE	Method of Lines as convolution plus differentiable ODE solver
Mesh-based variational learning	NeuFENet	Predict finite-element coefficients and train with FEM energy
Discretization-aware residual learning	NBM	Minimize residuals of a trusted numerical stencil at random collocation points
Hypernetwork operator learning	NPR	Map input function samples to parameters of a target neural field
Local operator discovery	Functional Convolution	Learn local stencil generators and solve by Picard iteration

NeuralPDE parameterizes the Method-of-Lines right-hand side with a CNN, yielding

$$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$

and then integrates this system with a differentiable ODE solver (Dulny et al., 2021). The same work also gives a higher-order formulation by augmenting the state with auxiliary variables, which proved useful on wave propagation. Its empirical profile is strongest on synthetic PDE benchmarks, while on real-world datasets such as Ocean Wave and WeatherBench it is competitive rather than uniformly dominant (Dulny et al., 2021).

NeuFENet moves much closer to classical finite elements. The solution is represented in a finite-element space,

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$

with the neural network predicting discrete FE coefficients rather than a coordinate-to-solution map. Training uses the Galerkin/Ritz energy

$J(u)=\frac12 B(u,u)-L(u),$

so Dirichlet conditions are imposed exactly on the discrete degrees of freedom and Neumann conditions enter naturally through the weak form (Khara et al., 2021). The central theoretical message is an error decomposition of the form

$$e_G \le \mathrm{Err}_\Theta+\mathrm{Err}_{\mathcal H}+O(h^{\alpha+1}),$$

which imports finite-element mesh convergence into the neural setting (Khara et al., 2021).

The Neural Bootstrapping Method takes a different route: it keeps the neural representation mesh-free, but computes residuals using an embedded discretization kernel. At each collocation point, the local residual is

$r_p=\|M^{-1}Au-M^{-1}b\|,$

where $A$ and $b$ come from a local finite discretization and $M^{-1}$ is a local preconditioner (Mistani et al., 2022). This makes the method “neuro-symbolic” in the paper’s sense because the PDE structure is enforced through a scientific-computing stencil rather than only through continuous autodiff residuals. Its demonstrated scope is elliptic interface problems with discontinuities (Mistani et al., 2022).

"Neural Partial Differential Equations with Functional Convolution" learns local stencil-generation rules rather than full solution maps. It represents a discretized operator as a shared local functional convolution, then solves nonlinear problems with Picard iteration and differentiates through the iterative solver with an adjoint-style linear-system calculation (Wu et al., 2023). The reported examples are notable for very small networks and training sets, with all examples trained with up to 8 data samples and within 325 network parameters (Wu et al., 2023).

3. Operator learning, reduced models, and explicit representations

A major line of work reframes NeuroPDE as operator learning. Neural Parameter Regression maps a discretized initial condition to the parameters of a target network:

$\theta = H_\Phi(u_0),$

and then evaluates the generated neural field $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$0 as the solution representation (Mundinger et al., 2024). Initial conditions are hard-coded via

$$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$1

so $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$2 exactly (Mundinger et al., 2024). Training is physics-informed through residual and boundary losses rather than paired solution trajectories, and low-rank factorization is used to reduce the target-network parameter count (Mundinger et al., 2024).

A related but conceptually distinct route is "Neural Control of Parametric Solutions for High-dimensional Evolution PDEs". There the reduced solution ansatz $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$3 defines a parameter manifold $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$4, and PDE evolution is pulled back to a control problem in parameter space:

$$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$5

The learned object is a neural control field $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$6, not the solution field itself (Gaby et al., 2023). New initial conditions are handled by fitting $$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$7 and evolving

$$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$8

For semilinear parabolic PDEs the paper gives an error bound

$$\frac{d\mathcal U}{dt}=\mathrm{CNN}_\theta(\mathcal U),\qquad \mathcal U(t_0)=\mathcal U_0,$$9

which separates initial encoding error and control-field approximation error (Gaby et al., 2023).

Neural SPDE extends this operator-learning picture to stochastic forcing. Based on the notion of mild solution of an SPDE, it introduces a neural architecture that learns solution operators of PDEs with possibly stochastic forcing from partially observed data, extends Neural CDEs and Neural Operators, and can be evaluated either by calling an ODE solver or by solving a fixed point problem (Salvi et al., 2021). The model is explicitly pathwise, with tasks of the form

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$0

and the reported stochastic KdV results show that including the realized noise path is essential (Salvi et al., 2021).

A plausible synthesis is that these methods replace the direct map from discretized input to discretized output by a more structured intermediate object: generated network parameters, latent parameter dynamics, or pathwise solution operators. This suggests a shift from “predict the field” to “learn the mechanism that generates the field.”

4. Stochasticity, path dependence, and non-Markovian extensions

Path dependence introduces an infinite-dimensional state variable, because the solution depends on the entire past trajectory rather than only the current state. PDGM extends the Deep Galerkin Method to this regime by combining an LSTM with a feed-forward network and minimizing a PPDE residual in path space (Saporito et al., 2020). Its functional derivatives are Dupire derivatives approximated by path perturbations:

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$1

and the training objective combines a PPDE residual term with a terminal-condition penalty (Saporito et al., 2020).

"A Neural RDE-based model for solving path-dependent PDEs" uses a different path representation. It reformulates the PPDE through functional Feynman–Kac as a conditional expectation problem, then uses an NRDE whose hidden dynamics are driven by truncated log-signatures of path segments (Fang et al., 2023). The model evolves

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$2

with $u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$3 built from log-signature features on each interval (Fang et al., 2023). For higher-dimensional problems the paper introduces EL-NRDE, which first embeds the path into a lower-dimensional signal before computing log-signatures. On heat- and Black–Scholes-type PPDE benchmarks, EL-NRDE consistently improves over EL-SigLSTM and uses less memory (Fang et al., 2023).

Neural SPDE belongs in the same extension family, but along the stochastic-forcing axis rather than the non-Markovian path-space axis. Its significance lies in making the learned object depend jointly on the initial condition and a realization of the driving noise, thereby moving beyond deterministic operator learning (Salvi et al., 2021).

These extensions correct a common misconception: neural PDE methods are not restricted to Markovian state PDEs on finite-dimensional spatial grids. In the surveyed literature they already cover PPDEs, SPDEs, and pathwise stochastic operator learning (Saporito et al., 2020, Fang et al., 2023, Salvi et al., 2021).

5. Constraints, weak forms, and numerical structure

Another major theme is how to incorporate classical numerical structure into neural PDE solvers. "An Unconstrained Formulation of Some Constrained Partial Differential Equations and its Application to Finite Neuron Methods" rewrites a Dirichlet elliptic problem as a saddle-point system and then applies an augmented Lagrangian Uzawa iteration at the PDE level, producing a sequence of unconstrained elliptic PDEs whose solutions converge to the constrained solution (Jia et al., 2024). The first iterate is precisely the standard penalty formulation,

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$4

but the paper shows that the second and later elements of the sequence are the ones that lead to the optimal $u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$5-norm error bound (Jia et al., 2024).

This perspective aligns closely with NeuFENet’s variational finite-element training and with NBM’s stencil-level residual minimization. Across these methods, neural approximation is embedded inside a structure inherited from weak formulations, saddle systems, or classical discretizations rather than replacing them wholesale (Khara et al., 2021, Mistani et al., 2022). A plausible implication is that a large fraction of NeuroPDE research is converging toward hybrid designs in which approximation flexibility is delegated to neural networks, but stability, constraints, and convergence remain organized by mature numerical analysis.

6. Interpretability, symbolic methods, and explicit formulas

Interpretability is a persistent concern. "Interpretable Neural PDE Solvers using Symbolic Frameworks" is a position paper arguing that the next step for neural PDE methods should be the integration of symbolic programming, especially symbolic regression and symbolic surrogate extraction, so that learned behavior can be rendered as human-readable mathematical structure (Lee, 2023). It distinguishes solution learning and operator learning, surveys PINNs, DeepONets, FNO, WNO, and PINO, and advocates three integration patterns: post hoc symbolic extraction from internal computations, neural-to-symbolic distillation of derivatives or dynamics, and symbolic metamodeling (Lee, 2023).

"Neuro-Symbolic Multitasking: A Unified Framework for Discovering Generalizable Solutions to PDE Families" makes this agenda algorithmic. NMIPS solves multiple related PDE tasks jointly in a multifactorial evolutionary search, uses a unified symbolic encoding space, and applies an affine transfer module parameterized by MLPs to move useful structure across tasks (Huang et al., 12 Feb 2026). Candidate solutions are evaluated by data loss, PDE residual, and initial/boundary-condition penalties, constants are refined by gradient-based optimization, and the final output is an explicit symbolic expression rather than a numerical field (Huang et al., 12 Feb 2026). Across six PDE families, the paper reports promising improvements over baselines, with up to a $u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$6 increase in accuracy while providing interpretable analytical solutions (Huang et al., 12 Feb 2026).

This strand highlights another misconception: interpretability in NeuroPDE need not mean feature attribution or saliency. In these papers it means operator-level or equation-level transparency—closed-form laws, explicit update rules, or symbolic surrogates (Lee, 2023, Huang et al., 12 Feb 2026).

7. Domain-specific instantiations and hardware realizations

NeuroPDE methods also appear as embedded components inside domain models rather than as general-purpose solvers. In "Hybrid PDE-Deep Neural Network Model for Calcium Dynamics in Neurons", the PDE backbone for cytosolic and ER calcium transport is retained, while a DNN replaces the Ryanodine receptor open-probability ODE subsystem through

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$7

and the learned $u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$8 is injected into the ER membrane flux (Gurung et al., 2024). The network does not solve the PDE; it learns a closure law inside a mechanistic PDE model. This is a paradigmatic hybrid mechanistic-learning design.

"Learning Mechanistic Subtypes of Neurodegeneration with a Physics-Informed Variational Autoencoder Mixture Model" goes further toward probabilistic mechanistic inference. BrainPhys embeds reaction-diffusion PDEs inside a VAE mixture model so that different subjects can be explained by different mechanistic components,

$u^h(x;\theta)=\sum_{i=1}^N \phi_i(x)U_i(\theta),$9

with subtype-conditioned latent diffusivity and reaction parameters inferred from sparse longitudinal PET data (Pinnawala et al., 18 Sep 2025). The decoder is a differentiable PDE solver rather than a PINN-style residual network. By contrast, "Learning Image Derived PDE-Phenotypes from fMRI Data" is a data-driven PDE-discovery workflow on a reduced latent representation of fMRI rather than a direct anatomical brain-space PDE; the paper itself is best interpreted as latent/reduced-order PDE discovery for biomarkers rather than direct neurobiological PDE modeling (Bica et al., 2024).

At the hardware extreme, "NeuroPDE: A Neuromorphic PDE Solver Based on Spintronic and Ferroelectric Devices" maps Monte Carlo random-walk PDE solving directly onto MTJ neurons and FTJ synapses (Fu et al., 7 Jul 2025). For diffusion benchmarks, the reported hardware variance relative to analytical solutions remains below $J(u)=\frac12 B(u,u)-L(u),$0, with speedup of $J(u)=\frac12 B(u,u)-L(u),$1 to $J(u)=\frac12 B(u,u)-L(u),$2 and energy advantage of $J(u)=\frac12 B(u,u)-L(u),$3 to $J(u)=\frac12 B(u,u)-L(u),$4 over advanced CMOS-based neuromorphic chips (Fu et al., 7 Jul 2025). Here “NeuroPDE” is literal neuromorphic hardware, not merely neural-network software.

Taken together, these application-specific and hardware papers broaden the meaning of NeuroPDE. The field includes not only neural surrogates for canonical PDE benchmarks, but also learned closures in biophysical models, probabilistic mechanistic decoders in neuroimaging, latent PDE-derived biomarkers, and stochastic hardware for random-walk solvers (Gurung et al., 2024, Pinnawala et al., 18 Sep 2025, Bica et al., 2024, Fu et al., 7 Jul 2025).