Compressible Neural Particle Method

• The paper demonstrates that cNPM extends neural particle methods by integrating pressure-density closure via the Tait equation for compressible viscous flow.
• It employs a physics-informed loss combining compressible Navier-Stokes equations and boundary conditions to ensure stability even with irregular particle distributions.
• Empirical results on dam-breaking scenarios show that cNPM delivers smooth, accurate fluid dynamics comparable to SPH while reducing instability issues.

The Compressible Neural Particle Method (cNPM) is a feed-forward neural network-based method that extends the original neural particle method to model compressible viscous fluid flows. It uses neural networks to calculate the velocity and pressure of fluid particles at the next time step, and the Tait equation to calculate the density to handle the compressibility. Its loss function is composed of the governing equations of compressible flow and the boundary conditions, which are free surface and solid boundary conditions. The method was demonstrated on a dam breaking problem and positioned as a stable alternative in settings where smoothed particle hydrodynamics (SPH) can become unstable depending on the distribution of particles (Shibukawa et al., 23 Aug 2025).

1. Origins, scope, and problem setting

cNPM emerged from the observation that particle methods are important in computational fluid dynamics, but are among the most difficult to implement and solve. In the formulation that motivated the method, SPH is identified as the most common particle method and as suitable for problem settings that involve large deformations, such as tsunamis and dam breaking. The same account states that SPH calculations can become unstable depending on the distribution of particles, whereas the neural particle method has high computational stability for various particle distributions and approximates velocity and pressure in a spatial domain using neural networks (Shibukawa et al., 23 Aug 2025).

Within that lineage, earlier neural particle methods had been extended to viscous flows but were limited to incompressible flows. cNPM is defined specifically as the extension that introduces compressibility into this framework. In this sense, it is neither a conventional SPH discretization nor a generic neural surrogate trained only from rollout data: it is a particle-based, physics-informed formulation in which the neural network is trained at each time step using the governing equations and boundary constraints (Shibukawa et al., 23 Aug 2025).

The method should also be distinguished from neighboring strands of research. Physics-informed neural networks for weakly compressible flows using the Galerkin-Boltzmann formulation reduce the state to a 6-equation first-order conservation law and use two distinct neural networks to manage scale disparity between equilibrium and non-equilibrium states (Aygun et al., 2024). NeuralMPM, by contrast, is a neural emulation framework inspired by the Material Point Method that interpolates particles to a fixed-size grid, updates the grid with image-to-image neural networks, and trains against ground-truth trajectories (Sharabi et al., 2024). These contrasts place cNPM in the narrower category of physics-informed neural particle solvers for compressible viscous flow rather than grid-based neural emulators or weakly compressible moment-based PINNs.

2. Governing equations and compressibility model

The governing equations enforced by cNPM are the compressible Navier-Stokes equations in Lagrangian, viscous form:

$$\frac{\partial \rho}{\partial t} = -\rho\, \nabla \cdot \bm{v}$$

$$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$

where $\bm{v}$ is velocity, $p$ is pressure, $\rho$ is density, $\mu$ is the dynamic viscosity coefficient, and $\bm{g}$ is gravity (Shibukawa et al., 23 Aug 2025).

Compressibility is introduced through the Tait equation:

$$\rho = \rho_0 \left( \frac{p}{B} + 1 \right)^{1/\gamma}$$

with

$B = \frac{\rho_0 c_s^2}{\gamma}$

and $\gamma$ chosen as $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$0 for water. In the reported implementation, the neural network outputs pressure, and density is then computed from this equation of state. The parameter $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$1 in the sound-speed definition controls how compressible the simulated flow is; smaller $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$2 corresponds to less compressibility. Choosing $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$3 and $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$4 recovers nearly incompressible behavior (Shibukawa et al., 23 Aug 2025).

The loss is decomposed as

$$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$5

where $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$6 is the error in the momentum equation, $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$7 is the error in mass conservation, $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$8 is the free surface boundary loss, and $$\frac{\partial \bm{v}}{\partial t} = - \frac{1}{\rho} \nabla p + \frac{1}{\rho} \nabla \cdot (\mu \nabla \bm{v}) + \bm{g}$$9 is the solid boundary Dirichlet loss (Shibukawa et al., 23 Aug 2025).

This formulation is significant because the method does not evolve density through a separate neural output channel. Instead, it uses a pressure output together with an equation of state and a continuity residual. This suggests that the reported notion of “compressible” is organized around pressure-density closure and mass conservation rather than a full thermodynamic closure with an explicit energy equation, a reading that is consistent with the paper’s stated future direction of adding true energy conservation equations (Shibukawa et al., 23 Aug 2025).

3. Neural approximation, time stepping, and differentiation

At the core of cNPM is a feed-forward neural network representing the mapping

$\bm{v}$0

for all fluid particles at each time step. The inputs are particle positions, and the outputs are the two velocity components and pressure. The network is trained at each time step to output the velocity and pressure for the next time step, based on current particle positions (Shibukawa et al., 23 Aug 2025).

The architecture reported in the paper consists of 4 layers, 50 hidden neurons per layer, the Mish function as activation, Adam as optimizer, and a learning rate of $\bm{v}$1 (Shibukawa et al., 23 Aug 2025). The parameter update strategy uses transfer learning: the network parameters at time $\bm{v}$2 are used to initialize the network at $\bm{v}$3. This is motivated by the expectation of small changes between adjacent timesteps and is reported to accelerate convergence during training at each time step (Shibukawa et al., 23 Aug 2025).

Temporal derivatives in the loss are handled using a Runge-Kutta scheme, propagating values backward from $\bm{v}$4 to $\bm{v}$5 (Shibukawa et al., 23 Aug 2025). The paper also states that viscous terms, specifically $\bm{v}$6, are computed using the neural network’s output and automatic differentiation, and that proper evaluation in a Lagrangian frame is accomplished following the Update Lagrangian formulation (Shibukawa et al., 23 Aug 2025).

These design choices align cNPM with the broader PINN literature but preserve its particle-based character. In weakly compressible Galerkin-Boltzmann PINNs, the principal architectural innovation is a two-network decomposition between equilibrium and non-equilibrium moments plus Fourier feature embeddings (Aygun et al., 2024). cNPM instead uses a single feed-forward network over particle coordinates and places the burden of compressibility on the Tait equation, the continuity residual, and boundary losses. The result is a more direct particle formulation centered on next-step field prediction.

4. Boundary conditions, free surfaces, and wall treatment

Boundary handling is integral to the cNPM formulation. At the free surface, pressure is required to equal atmospheric pressure, typically set to zero for water in atmospheric contact, and this condition is enforced through the free-surface loss term $\bm{v}$7 (Shibukawa et al., 23 Aug 2025). At solid walls, the no-penetration or no-slip condition is enforced by requiring velocity to be zero at solid boundaries, which is encoded in the solid-boundary loss $\bm{v}$8 (Shibukawa et al., 23 Aug 2025).

To prevent wall penetration, the method additionally applies a repulsive penalty force to particles near the solid boundary. The penalty depends on signed distance from the wall, the wall normal, and a stiffness coefficient (Shibukawa et al., 23 Aug 2025). This supplements the loss-based enforcement of no-slip behavior with an explicit geometric barrier mechanism.

An important implementation detail is that the set of particles identified as free surface or wall remains fixed in this method and is not updated at runtime (Shibukawa et al., 23 Aug 2025). That detail has direct implications for interpretation. A plausible implication is that the reported results are strongest in settings where these boundary-designated particle sets can be specified in advance and remain adequate throughout the evolution. The paper does not present a runtime boundary reclassification procedure.

This boundary treatment differs both from SPH-style particle-boundary coupling and from grid-based neural emulators. In adaptive particle refinement for compressible SPH, resolution management is performed by explicit splitting and merging in user-specified regions, with conservation of mass and careful treatment of momentum and energy (Nealon et al., 2024). cNPM does not rely on such refinement machinery in its reported form; its emphasis is on stable neural field approximation over a particle set with fixed surface and wall labels.

5. Empirical behavior on dam breaking and comparison with SPH

The numerical study reported for cNPM is a dam breaking problem with water (Shibukawa et al., 23 Aug 2025). The paper evaluates both stability to particle distribution and agreement with an SPH baseline.

For particle-distribution sensitivity, cNPM is tested on both uniform-grid and randomized initial particle positions. In both cases, it yields smooth, stable density and pressure fields. Even with local particle voids or clusters, no spurious spikes or noise appear, the field remains physically coherent, and the free surface is smooth (Shibukawa et al., 23 Aug 2025). This is presented in direct contrast with conventional SPH, which is described as known to develop significant noise, density or pressure oscillations, or even numerical divergence under irregular particle arrangements (Shibukawa et al., 23 Aug 2025).

For quantitative comparison, cNPM and implicit SPH (iSPH) are compared through pressure histories at several wall locations. Both predict the same timing and magnitude for pressure peaks and decay, although cNPM sometimes shows slightly higher peaks and otherwise very similar time traces (Shibukawa et al., 23 Aug 2025). This is the paper’s primary evidence that cNPM matches SPH quantitatively in pressure and global dynamics when both methods are stable.

The paper also reports a compressibility parametric study based on the Tait-equation parameter $\bm{v}$9. For nearly incompressible settings with $p$0, density remains essentially constant. For highly compressible settings with $p$1, density variations in high-pressure regions grow, but the overall macroscopic flow and dam front are not drastically altered (Shibukawa et al., 23 Aug 2025). This result indicates that the reported benchmark is sensitive to density variation while remaining dominated by the same large-scale surge kinematics across the tested compressibility range.

Finally, transfer learning is reported to reduce the initial loss and to attain convergence with fewer epochs at each time step, thereby accelerating computation without loss in accuracy (Shibukawa et al., 23 Aug 2025). The paper therefore presents efficiency not as a global offline training claim, but as reduced per-step training burden in an online or sequentially retrained solver.

6. Position in the literature, misconceptions, and prospective extensions

cNPM belongs to a broader family of neural particle methods, but its specific combination of compressible viscous flow, equation-of-state closure, and timestep-wise physics-informed training makes it distinct. Deep JKO develops neural network-based implicit particle methods for high-dimensional Wasserstein-type gradient flows, with velocity fields approximated by neural networks and density tracked through continuous normalizing flow machinery (Lee et al., 2023). Neural Particle Automata use differentiable SPH operators and shared local neural rules to learn self-organizing particle dynamics with continuous positions and internal states (Kim et al., 22 Jan 2026). These methods share Lagrangian and neural elements with cNPM, but they target different governing principles and objectives.

A common misconception is to treat cNPM as a neuralized version of SPH. The paper’s reported mechanism is different: the network predicts velocity and pressure; density is obtained from the Tait equation; the loss is built from mass and momentum equations plus boundary conditions; and viscous terms are evaluated through automatic differentiation in an Update Lagrangian formulation (Shibukawa et al., 23 Aug 2025). This is closer to a particle-based PINN than to a kernel-summation SPH scheme.

A second misconception is to interpret the method as a general compressible-flow solver in the strongest thermodynamic sense. The paper explicitly identifies future directions involving extension to higher compressibility, more general equations of state such as Tillotson, and adding true energy conservation equations (Shibukawa et al., 23 Aug 2025). This suggests that the current formulation is best read as a demonstrated method for compressible viscous free-surface flow with Tait-equation closure, rather than as a complete high-Mach or general-EOS compressible framework.

In the non-neural particle-method literature, recent work on adaptive particle refinement for compressible SPH reports splitting and merging, multiple dynamic refinement regions, errors in the mass accreted onto sinks of less than 9 percent, and speed ups of 1.07–6.62 times for the examples shown (Nealon et al., 2024). A plausible implication is that such refinement infrastructure could complement cNPM in future hybrid schemes where regions of interest trigger local resolution changes, although the adaptive-refinement paper explicitly does not mention neural or learned particle systems (Nealon et al., 2024).

Taken together, the reported evidence supports a precise characterization of cNPM: a feed-forward, physics-informed neural particle formulation for compressible viscous fluid flow in which pressure and velocity are predicted at the next time step, density is obtained from the Tait equation, free-surface and solid-wall conditions are imposed through loss terms, and the principal demonstrated advantage is stable behavior under irregular particle distributions in a dam-breaking benchmark (Shibukawa et al., 23 Aug 2025).