Neural Particle Method

• Neural Particle Method is a family of techniques that combines particle-based representations with neural parameterizations to learn and propagate uncertainty.
• It is applied across sequence modeling, PDEs, and fluid dynamics, replacing hand-crafted operators with differentiable, learned components.
• The method achieves improved accuracy and efficiency, enabling robust continuous-time filtering, physics-informed simulation, and scalable high-dimensional inference.

Neural particle method denotes a family of approaches that combine particle-based representations with neural parameterizations and optimization. Across the literature, the particles may represent weighted latent hypotheses, unweighted posterior samples, moving material points, empirical measures of stochastic systems, or particles carrying internal states, while the neural component may parameterize a latent transition, an observation likelihood, a score function, a characteristic flow map, a field interpolator, a local update rule, or a transport map between measures (Ma et al., 2019, Kutschireiter et al., 2015, Wessels et al., 2020, Kim et al., 11 Mar 2026, Kim et al., 15 Oct 2025, Kim et al., 22 Jan 2026, Wang et al., 2021). This suggests that the phrase functions less as the name of a single standardized algorithm than as an umbrella designation for methods that keep the sample-based, meshfree, or empirical-measure viewpoint of particle methods while replacing hand-crafted operators with learned and differentiable components.

1. Terminological scope and historical usage

A notable early formulation is the Neural Particle Filter, which introduced a weight-less particle filter interpretable as the dynamics of a recurrent neural network performing nonlinear Bayesian filtering in continuous time (Kutschireiter et al., 2015). In sequence modeling, Particle Filter Recurrent Neural Networks replaced a single deterministic hidden state with a particle approximation of a latent-state distribution and embedded a differentiable particle filter into gated RNNs (Ma et al., 2019). In computational fluid dynamics, the expression “Neural Particle Method” was used more specifically for an Updated Lagrangian, meshfree, physics-informed neural network for incompressible free-surface Euler flow (Wessels et al., 2020), and was later extended to compressible viscous flows through the compressible Neural Particle Method, which adds a continuity equation and a Tait equation of state (Shibukawa et al., 23 Aug 2025).

More recent work broadened the scope further. “A Physics-Informed, Global-in-Time Neural Particle Method for the Spatially Homogeneous Landau Equation” parameterizes both a score field and a characteristic flow map with neural networks and enforces Landau dynamics through a continuous-time residual along trajectories (Kim et al., 11 Mar 2026). “Solving McKean-Vlasov Equation by deep learning particle method” treats particle trajectories as neural functionals of Brownian paths and matches pseudo SDE coefficients to the original MV-SDE coefficients through a PINN-style loss independent of the Euler step size (Li et al., 1 Jan 2025). Other variants include a neural score-based particle method for the Vlasov–Maxwell–Landau system (Ilin et al., 26 Mar 2026), a neural stochastic interacting particle-field algorithm for chemotaxis with CNN interpolation (Kim et al., 15 Oct 2025), and Neural Particle Automata, which generalize Neural Cellular Automata from lattices to moving particles with SPH-based perception (Kim et al., 22 Jan 2026).

2. Core representations and algorithmic primitives

A recurrent structural feature is the replacement of a single state by an empirical distribution. In PF-RNNs, the latent belief is approximated by

$$b(h_t) \approx \sum_{i=1}^{K} w_t^i \, \delta(h_t - h_t^i),$$

with prediction formed from the weighted mean

$\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$

The particles share parameters, so increasing $K$ increases computation but not the number of trainable weights (Ma et al., 2019). In the Neural Particle Filter, the posterior is represented by equally weighted samples,

$$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$

so the particle cloud itself, rather than a weighted importance representation, carries the posterior (Kutschireiter et al., 2015).

In scientific computing, the particle representation is often literal. NSIPF approximates the density by

$$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$

while a CNN supplies a smooth chemoattractant field used to compute the drift of the particles (Kim et al., 15 Oct 2025). DeepParticle uses empirical measures

$$\mu=\frac{1}{N}\sum_{i=1}^{N}\delta_{x_i}, \qquad \nu=\frac{1}{N}\sum_{j=1}^{N}\delta_{y_j},$$

and learns a neural transport $f_\theta$ by minimizing a discrete Wasserstein distance between the pushed-forward source particles and target particles generated by an interacting particle method (Wang et al., 2021). Neural Particle Automata attach to each particle a continuous position $\mathbf{x}_i$, a mass $m_i$, and an internal state $\mathbf{S}_i$, and update both state and position using a shared local MLP driven by SPH features (Kim et al., 22 Jan 2026).

These examples indicate that “particle” can refer either to hypotheses in a latent posterior, samples from a measure, or meshfree material points. The commonality lies in representing uncertainty, geometry, or solution structure by discrete samples and then learning the operators that propagate or interpret those samples.

3. Sequential inference and latent-state modeling

PF-RNNs instantiate the neural particle method in sequence learning by turning the hidden memory of an LSTM or GRU into a learned particle filter. For each particle, the transition is stochastic,

$\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$0

weights are updated in Bayes-like form,

$\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$1

and resampling is replaced by soft resampling with

$\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$2

The method is fully differentiable through reparameterized transition noise and soft resampling, and training combines a prediction loss with an ELBO-style term using $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$3. In experiments, PF-RNNs outperform corresponding standard gated RNNs on a synthetic robot localization dataset and 10 real-world sequence prediction datasets, outperform Bayesian LSTM/GRU in 16 of 20 dataset-model combinations, and substantially outperform ensembles with the same number of copies as particles; the reported design choices include $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$4 particles and $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$5 (Ma et al., 2019).

The Neural Particle Filter addresses continuous-time nonlinear Bayesian filtering without importance weights. Its particle dynamics are

$\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$6

where the innovation term $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$7 acts as a prediction error. The method is interpreted as a recurrent network of filtering neurons receiving feed-forward input from novelty neurons, and it supports online maximum-likelihood learning of model parameters. The reported numerical performance is adequate for both filtering and identification, and the weightless construction is claimed to alleviate the curse of dimensionality and to outperform conventional weighted particle filters in higher dimensions for a limited number of particles (Kutschireiter et al., 2015).

In Bayesian neural networks, a related particle idea appears in function-space particle optimization. Instead of moving particles in weight space, the method optimizes particles in the space of regression functions, using kernels on $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$8 and backpropagating the function-space updates to weights. The reported result is that function-space particle optimization avoids the failure mode in which diverse weight vectors correspond to nearly identical functions, and it outperforms strong baselines in regression, adversarial robustness, and reinforcement learning; on MNIST it reports a test error of $\bar{h}_t = \sum_{i=1}^{K} w_t^i h_t^i.$9 (Wang et al., 2019). Although this setting is not a particle method in the meshfree PDE sense, it fits the same sample-based design pattern.

4. Physics-informed particle solvers for PDEs and SDEs

In fluid mechanics, the original Neural Particle Method is an Updated Lagrangian PINN for incompressible, inviscid Euler flow with free surfaces. A feed-forward network takes particle positions at time $K$0 and outputs velocity stages, next-step velocity, and pressure stages, while the loss enforces implicit Runge–Kutta consistency, incompressibility, and boundary conditions. Spatial derivatives are obtained by automatic differentiation, and particles are then moved by the IRK position update. The method is explicitly meshfree, remains stable and accurate even if the location of discretization points is highly irregular, and avoids the specialized incompressibility stabilization used in many classical formulations (Wessels et al., 2020). The compressible Neural Particle Method extends this framework to compressible viscous flow by predicting velocity and pressure at the next time step, computing density through the Tait equation, and training against mass conservation, momentum conservation, and free-surface and solid-wall losses. In a dam-breaking problem it is reported to accurately solve compressible viscous fluid flow that was difficult to solve with SPH, including under random initial particle distributions; the main configuration uses 64,800 total simulation particles and 800 training particles per step, and transfer learning accelerates convergence between time steps (Shibukawa et al., 23 Aug 2025).

For kinetic equations, PINN–PM introduces a global-in-time Lagrangian interacting-particle solver for the spatially homogeneous Landau equation. It parameterizes both the score $K$1 and the flow map $K$2, and enforces the characteristic ODE through the residual

$K$3

Training combines a physics loss with implicit score matching evaluated along learned trajectories. Because time is an input to the networks, the method removes time-discretization error and can be queried at arbitrary times without sequential integration. The analysis yields trajectory and density error bounds in an $K$4 and Wasserstein framework, with error decomposed into score approximation error, empirical particle approximation error, and physics residual. On BKW benchmarks and reference-free cases, the method reports stable transport, preservation of macroscopic invariants, and competitive or improved accuracy relative to time-stepping score-based particle and blob methods while using significantly fewer particles; the experiments use 1,000 particles, whereas SBP uses 14,400–64,000 (Kim et al., 11 Mar 2026).

The Vlasov–Maxwell–Landau score-based transport model keeps the deterministic particle/PIC structure but replaces the blob score estimator by an on-the-fly neural score $K$5 trained through implicit score matching,

$K$6

The resulting approximate collision operator preserves momentum and kinetic energy and dissipates an estimated entropy. On Landau damping, two-stream instability, and Weibel instability, the method is more accurate than the blob method, achieves correct long-time relaxation to Maxwellian equilibrium where the blob method fails, and delivers $K$7 faster runtime with $K$8 lower peak memory (Ilin et al., 26 Mar 2026).

For stochastic interacting systems, the deep learning particle method for MV-SDEs represents particle trajectories as neural functionals $K$9, derives a pseudo SDE by Itô calculus, and minimizes the discrepancy between pseudo coefficients and the true MV-SDE coefficients. The approximation error is controlled by $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$0 and $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$1, rather than an Euler step size $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$2, and the authors emphasize that the error is independent of the time step size (Li et al., 1 Jan 2025). NSIPF for a parabolic–hyperbolic Keller–Segel system combines a stochastic particle approximation of density with a CNN approximation of the chemoattractant field. The particle update is

$$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$3

and the reported convergence rates are $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$4 and $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$5. At grid sizes $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$6, the reported runtimes for NSIPF are $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$7 s, $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$8 s, and $$p(x_t \mid Y_t) \approx \frac{1}{N}\sum_{k=1}^N \delta(x-z_t^{(k)}),$$9 s, versus $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$0 s, $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$1 s, and $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$2 s for classical SIPF with spline interpolation (Kim et al., 15 Oct 2025).

5. Broader extensions of the neural-particle paradigm

Neural Particle Automata recast Neural Cellular Automata in a Lagrangian setting. Each particle carries position, mass, and internal state, local perception is built from SPH operators such as density, smoothed state, and state gradients, and a shared two-layer MLP produces increments $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$3 and $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$4. The method is trained with rollout-based losses and exhibits characteristic NCA behaviors, including robustness, self-regeneration, persistent flows, and asynchronous stability. Reported tasks include morphogenesis, particle-based texture synthesis, and point-cloud classification, with a PointMNIST test accuracy of $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$5 (Kim et al., 22 Jan 2026).

DeepParticle addresses invariant-measure learning by training a parameter-dependent deep network $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$6 to map a simple source distribution to invariant measures generated by an interacting particle method. The discrete Wasserstein objective is

$$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$7

with the optimal transport subproblem solved by an iterative divide-and-conquer mini-batch interior-point algorithm. In 2D and 3D chaotic-flow front-speed problems, the learned map provides warm starts that accelerate interacting particle computation of invariant measures and principal eigenvalues by $$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$8–$$\rho_t(x) \approx \frac{M_0}{P}\sum_{p=1}^{P}\delta(x-X_t^p),$$9 (Wang et al., 2021).

In solid mechanics, NN-RKPM enriches a reproducing kernel particle approximation with blockwise neural localization functions:

$$\mu=\frac{1}{N}\sum_{i=1}^{N}\delta_{x_i}, \qquad \nu=\frac{1}{N}\sum_{j=1}^{N}\delta_{y_j},$$0

The NN part captures the location, orientation, and shape of strain-localization zones, while the RKPM part approximates the smooth background field. A basic four-kernel NN block can capture a triple junction or a quadruple junction topological pattern, and a regularization of the NN approximation is introduced to obtain discretization-independent material responses. In a 1D benchmark, the paper reports that 21 RK nodes plus 4 NN blocks, for a total of 57 DOFs, match an exact localized solution, whereas a pure RKPM discretization needs 801 DOFs for comparable resolution (Baek et al., 2022).

6. Numerical properties, theoretical structure, and limitations

Across the literature, neural particle methods are used to address three recurring numerical difficulties: multimodal latent uncertainty, stiff or high-dimensional transport, and the need to resolve localized structure without globally refining a mesh or grid.