Applied Ocean Research 154 (2025) 104424

Contents lists available at ScienceDirect

Applied Ocean Research

journal homepage: www.elsevier.com/locate/apor

Step-by-step enhancement of a graph neural network-based surrogate

model for Lagrangian fluid simulations with flexible time step sizes
Takefumi Higaki a,* , Yuki Tanabe a , Hirotada Hashimoto a , Takahito Iida b
a
Department of Marine System Engineering, Graduate School of Engineering, Osaka Metropolitan University, 1-1 Gakuen-cho, Naka-ku, Sakai, Osaka, 599-8531, Japan
b
Department of Naval Architecture and Ocean Engineering, Graduate School of Engineering, Osaka University, 2-1 Yamadaoka, Suita, Osaka, 565-0871, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: Machine learning (ML) has the potential to accelerate particle-based Lagrangian fluid simulations. However,
Particle-based surrogate model previous studies have not sufficiently validated the generalization performance of ML-based surrogate models
Graph network-based simulators across different fluid phenomena. In addition, most surrogate models have been verified under the Courant-
Moving particle semi-implicit
Friedrichs-Lewy (CFL) condition, similar to traditional CFD methods, which limits their ability to drastically
Pressure poisson equation
Courant-Friedrichs-Lewy condition
reduce computation time. Furthermore, many existing studies focus solely on the final methodology and results,
Violent free-surface flow without clarifying the effect of different feature settings on the outcomes. In this paper, we present a particle-
based surrogate model that can handle flexible time step sizes and different fluid phenomena, along with its
gradual improvement. Our proposed method, a graph network-based simulator with wall boundary nodes and
pressure estimation (GNS-WP), was first evaluated step-by-step on dam-break and sloshing tests. Then, we
demonstrated the model’s convergence properties with respect to time step and particle sizes. Finally, the GNS-
WP model, trained in a sloshing scenario, was validated on three benchmark tests with different wall boundaries
and flow dynamics. The proposed method achieved accuracy equal or better than the particle method used as the
teacher, with substantial computational speedups.

1. Introduction (2018), Mrowca et al. (2018), and Li et al. (2018) have dedicated efforts
to the fast computation of Lagrangian fluid simulations. However, these
Fluid simulation is essential in ocean and coastal engineering. Par- studies focused on replicating fluid-like behavior in simulation envi-
ticle methods, such as smoothed particle hydrodynamics (SPH; Gingold ronments rather than on accuracy.
and Monaghan, 1977), moving particle semi-implicit (MPS; Koshizuka Afterwards, studies on the surrogate models for SPH have made
and Oka, 1996), or incompressible SPH (ISPH; Cummins and Rudman, progress, with a gradual improvement in their accuracy. Zhang et al.
1999; Lo and Shao, 2002), have been widely used due to their (2020) used the results of SPH as training data to perform Lagrangian
Lagrangian and mesh-free nature, of which applications are compre- fluid analysis with deep neural networks (DNN). Ummenhofer et al.
hensively summarized by Luo et al. (2021). Despite their advantages, (2019) presented surrogate solvers for both NVIDIA FleX and SPH by
computational cost remains a major concern in particle methods. In using continuous convolution. Sanchez-Gonzalez et al. (2020) intro-
recent years, there has been a growing need to simulate real-world fluids duced graph network-based simulators (GNS), which made graph neural
in every phase of engineering, including design, development, networks (GNN) to learn the motion of fluid particles from SPH data.
manufacturing, verification, operation, and visualization. Therefore, Prantl et al. (2022) proposed guaranteeing the conservation of mo-
computation time should be further reduced. mentum by antisymmetric continuous convolution layers. However,
Over the last decade, researchers have presented machine-learning- these previous studies have not aimed to estimate pressure, which is one
based surrogate models that estimate fluid dynamics with a smaller of the most important factors in understanding fluid mechanics and its
computational cost. Ladický et al. (2015) pioneered using machine interaction with structures.
learning to substitute particle methods. They reformulated the Recent trends have focused on the pressure of incompressible fluids.
Navier-Stokes equation as a regression problem and estimated acceler- Li and Farimani (2022) presented fluid graph networks (FGN), which
ations by applying random forest. Subsequently, Schenck and Fox achieved both speed and accuracy by using MPS as training data and by

* Corresponding author.
E-mail address: [email protected] (T. Higaki).

https://doi.org/10.1016/j.apor.2025.104424
Received 22 July 2024; Received in revised form 15 December 2024; Accepted 7 January 2025
Available online 15 January 2025
0141-1187/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 1. Overview of GNS. In GNS, a node represents a fluid particle, and an edge represents the interaction between particles. The nodes within the range of the
connectivity radius RG are regarded as neighbors, in the same manner as MPS. First, edges are constructed among the neighboring nodes at every time step k. Second,
node and edge features are computed. Third, the encoders generate a latent graph from the features. Fourth, the processor updates the latent graph in the framework
of message passing, which is conducted M times. Fifth, the decoder predicts desired values (in the original GNS, the accelerations are estimated). Last, each network is
updated to minimize the loss function, i.e., the error from the target values obtained by particle methods.

Fig. 2. Comparison of the treatment of wall boundary between GNS and GNS-W.

employing node- and edge-specialized GNNs. Their work was the first to across different fluid phenomena. In addition, Zhang et al. (2023) and
predict hydrodynamic pressure by a particle-based surrogate model, as (2024a) have verified the convergence property using the constant
far as the authors’ knowledge. Bai et al. (2022) introduced a physically Courant number, which means fixing the ratio of time step size to par-
consistent surrogate model that required no training data, by embedding ticle size. Their verification was limited under the
boundary conditions and mass/momentum conservation into the loss Courant-Friedrichs-Lewy (CFL) condition like traditional CFD methods,
functions. Yao et al. (2023) accelerated solving pressure Poisson equa- which can hinder efforts to drastically reduce computation time.
tion (PPE) in MPS by using DNN. Nevertheless, none of these previous Moreover, most of the aforementioned studies have presented only the
studies have confirmed whether the estimated pressure reproduced the final methodology and results, without clarifying the effect of different
actual phenomena. feature settings on the results. Because the calculation results likely
Latest studies have investigated physical consistency as well as the depend on the feature settings, it is important to reveal which feature is
accuracy of pressure estimation. Zhang et al. (2023) achieved fast and essential for surrogate models to reproduce fluids.
accurate computation by training a convolutional neural network (CNN) Therefore, this paper presents a particle-based surrogate model
to solve the PPE in ISPH. They verified the fluid motions, pressure, and applicable to larger time step sizes and different fluid phenomena, along
the conservation of mass/volume and energy with several benchmark with its step-by-step enhancement. Our proposed methods are based on
tests. Their subsequent studies (Zhang et al., 2024a, b) replaced CNN GNS (Sanchez-Gonzalez et al., 2020) and use the computation results of
with GNN, which can treat unstructured data and thus are more suitable MPS developed by Iida and Yokoyama (2022b). In this study, we
for particle-based simulations. Although these papers have demon- introduced three enhanced versions of GNS: GNS with wall boundary
strated various applications, such as the interaction of a propagating nodes (GNS-W), GNS with pressure estimation (GNS-P), and GNS with
wave and a floating body, they prepared separate training data for each wall boundary nodes and pressure estimation (GNS-WP). First, we
specific purpose; for example, they trained CNN/GNN using showed that pressure estimation is important for accurate prediction of
dam-breaking data to simulate dam break with different boundaries. fluids by comparing these four models against the dam-break tests
None of the previous studies have validated generalization performance conducted by Martin and Moyce (1952) and Lobovský et al. (2014).

2
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 3. Schematics of the problem settings in (a) dam-break, (b) sloshing, (c) hydrostatic, and (d) free-oscillation-in-a-triangular-tank tests.

2. Methods
Table 1
Condition of each dataset. For training in the dam-break simulations, nine 2.1. Moving particle semi-implicit (MPS)
datasets (Datasets A4‒12) were used, whereas for training in the sloshing, one
dataset (Dataset B1) was used. The governing equations of fluid are as follows:
Dataset XF [m] YF [m] XW [m] YW [m] Ξ [m] T [s]
1 Dρ
A1 (for test; Martin and 0.30 0.60 1.61 1.00 + ∇⋅u = 0 (1)
– – ρ Dt
Moyce, 1952)
A2 (for test; Lobovský 0.60 0.60 1.61 1.00
Du 1
– –
et al., 2014) = − ∇p + ν∇2 u + g (2)
A3 (for test; Lobovský 0.60 0.30 1.61 1.00 – – Dt ρ
et al., 2014)
A4 (for training) 0.60 0.12 1.61 1.00 – – where ρ, p, ν, u, and g denote the fluid density, pressure, kinematic
A5 (for training) 0.68 0.60 1.61 1.00 – – viscosity, velocity, and gravitational acceleration, respectively. It is
A6 (for training) 0.68 0.30 1.61 1.00
common to model the interaction between particles i and j with the
– –
A7 (for training) 0.68 0.12 1.61 1.00 – –
A8 (for training) 0.60 0.60 1.18 1.00 – – following Wendland kernel (Wendland, 1995),
A9 (for training) 0.60 0.30 1.18 1.00 – – ⎧(
⎨ 1 − rij rij ) (
)4 ( )
A10 (for training) 0.60 0.12 1.18 1.00 – – ⎪
1+4 rij < R
A11 (for training) 0.68 0.60 1.18 1.00 – – wij = R R (3)
A12 (for training) 0.68 0.30 1.18 1.00 ⎪ ( )
– – ⎩ 0 rij ≥ R
A13 (for test; Hu and 0.68 0.12 1.18 1.00 – –
Kashiwagi 2004)
B1 (for training) 0.60 0.12 0.60 0.30 0.05 1.3 Here, R is called effective radius, within which the interaction between
B2 (for test; Kishev 0.60 0.12 0.60 0.30 0.05 1.5 particles is considered. rij represents the distance between particles i and
et al., 2006) ⃒ ⃒
C1 (for test) 0.20 0.10 0.20 – – – j, given by rij ≡ ⃒xj − xi ⃒ (xi is the position vector of particle i).
C2 (for test) 0.20 0.20 0.20 – – – Based on the projection method (Chorin, 1968), the calculation of
D1 (for test; Lamb, 0.28 0.14 – – – – MPS is divided into two steps: predictor and corrector steps. At the
1932)
predictor step, the intermediate velocity of particle i is calculated as
( )
u*i = uki + ν∇2 uki + g Δt (4)
Second, we confirmed the wall boundary nodes are essential to treat
moving wall boundary conditions by testing on the sloshing problem where k and Δt denote the present time step and time step size. Then, the
presented by Kishev et al. (2006). Third, we demonstrated that GNS-WP following PPE is solved at the intermediate positions and velocities.
achieved to reproduce the sloshing with sufficient accuracy even with
the time step size ten times larger than that of the training data. Fourth, 1 2 k+1 1
− ∇ pi = − ∇⋅u*i (5)
we examined the convergence property of GNS-WP with respect to the ρ Δt
particle size. Finally, we demonstrated that the proposed method,
trained in a sloshing scenario, can be applied to three different problems: Finally, the velocities and positions are updated by pressure gradient at
the dam-break (Hu and Kashiwagi, 2004), hydrostatic, and the corrector step.
free-oscillation-in-a-triangular-tank (Lamb, 1932) tests. 1
uk+1
i = u*i − ∇pk+1
i Δt (6)
ρ
( )
xk+1
i = x*i + uk+1
i − u*i Δt (7)

3
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 4. Time histories of the (a) horizontal position of the water front and (b) height of the residual water column in the test case of A1. The numerical results
computed by MPS, GNS, GNS-W, GNS-P, and GNS-WP are compared with the experiment data (Martin and Moyce, 1952) and the ISPH-based surrogate model
(ISPH_ML; Zhang et al., 2023).

To generate training dataset for our proposed surrogate models, this MPS, the nodes within the range of the connectivity radius RG are
study adopted the state-of-the-art MPS solver presented by Iida and regarded as neighbors, and edges are constructed among the neigh-
Yokoyama (2022b). The MPS solver modeled fixed or moving wall boring nodes at every time step.
boundaries by using wall particles as well as fluid particles and assigning Node and edge features are then computed. The feature of node i at
arbitrary velocities to the wall particles. The free surface was detected time step k was defined as follows:
by the parachute detection (PD; Marrone et al., 2010) and free surface [ ]
assessment by time tracing of free surface particles (FATT; Iida and nki ≡ ũki , u
̃k−i 1 , ⋯u ̃k
̃k−i 4 , Δ (8)
wall,i
Yokoyama, 2022a). As a free surface boundary condition, zero pressure
[ k k]
was imposed on the space potential particles (SPP; Tsuruta et al., 2015). uki u v
̃ki ≡
where u = i, i ,
To solve PPE with high accuracy, the resolution-free second-order dif- U U U
ferential source term (RF-SDS; Iida and Yokoyama, 2022b) was [ ( k ) ( k )]
k Δxwall,i Δywall,i
employed with the error compensating source (ECS; Khayyer and Gotoh, Δwall,i ≡ min
̃ , 1 , min ,1
2011). Higher-order Laplacian (HL; Khayyer and Gotoh, 2010) was also RG RG
used in PPE, whereas standard Laplacian was used in the viscous term. Here, uki is the velocity vector of node i at time step k. The tilde in-
The pressure gradient was enhanced by the combination of the dicates the variable is normalized; U is the representative speed for
higher-order gradient (HG; Iribe and Nakaza, 2011; Khayyer and Gotoh, normalization. Following the original GNS (Sanchez-Gonzalez et al.,
2011), dynamic stabilizer (DS; Tsuruta et al., 2013), and stable pressure 2020), the nondimensional velocities at the current and previous
gradient for free surface domains (Duan et al., 2017). The detailed ex- ̃ k , rep-
four-time steps (five steps in total) are given. The last term, Δ wall,i
planations are provided in the original paper (Iida and Yokoyama,
2022b). resents the wall boundary condition. Δxkwall,i and Δykwall,i are the hori-
zontal and vertical distances toward the nearest wall, which are
normalized with the connectivity radius RG and then clipped within the
2.2. Graph network-based simulators (GNS)
range [0, 1]. The feature of edge ij at time step k was also defined in line
with the original GNS,
This study employed GNS (Sanchez-Gonzalez et al., 2020), of which
overview is described in Fig. 1. In GNS, a node represents a fluid particle,
and an edge represents the interaction between particles. Similarly to

4
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

〈 〉0 ( )
ekij = fenc(e) ekij (11)

where fenc(n) and fenc(e) are NNs for encoding the node and edge features,
respectively. Next, the processor updates the latent graph as follows:
( )
〈 k 〉m+1 〈 〉m ∑ (〈 k 〉m 〈 k 〉m 〈 k 〉m )
ni = fupd nki , fmsg ni , nj , eij (12)
j

This procedure is called message passing, where the message function

fmsg aggregates the information on the vicinity of node i, and the update
function fupd updates the node status. Note that, in semi-implicit
methods like MPS, PPE is solved across the entire field to obtain glob-
ally consistent solutions. To consider not only local but global in-
teractions, message passing procedure is repeated M times. Next, the
decoder predicts a target value at time step k + 1.
(〈 〉M )
zk+1
i = fdec nki (13)

By multiplying the predicted value zk+1

i by the representative value A,
the acceleration of each node can be obtained.
[ k+1 k+1 ]
u̇i , v̇i = Azk+1
i
(14)

The objective of GNS is to minimize the following loss,

N N (
1 ∑ ( k+1 )2 1 ∑ ( k+1 k+1 )2 ( k+1 )2
)
L≡ zi − zk+1 = u̇i − u̇i + v̇k+1 − v̇i
N i i
NA2 i i

(15)

k+1 k+1
where N is the number of all nodes. Here, u̇ i and v̇ i are the ground-
truth accelerations of particle i at time step k + 1 obtained by particle
methods. The definitions above are consistent with the original GNS
(Sanchez-Gonzalez et al., 2020).

2.3. GNS with wall boundary nodes (GNS-W)

In the original GNS (Sanchez-Gonzalez et al., 2020), the wall

boundary condition was simply considered with the distances from the
nearest wall, as shown in Fig. 2. However, this is not appropriate if the
wall is moving or has complex geometry. To solve this issue, we intro-
duced wall boundary nodes, in the same manner as particle methods.
Instead of the distances from the nearest wall boundary, the particle type
qi was newly implemented in the node feature as follows:
[ k k− 1 ]
nki ≡ ũi , u ̃k−i 4 , qi
̃i , ⋯u (16)
{
1 if node i is fluid
where qi ≡
Fig. 5. Comparison of the snapshots in the test cases of (a) A2 and (b) A3. In 0 if node i is wall
MPS, GNS-P, and GNS-WP, particles or nodes are colored by their pressure.
Since this study targeted the problems with fixed or forced oscillating
[ ] walls, the wall boundary nodes were excluded when computing the loss,
xkj − xki ykj − yki rijk
ekij ≡ , , (9) NF (
RG RG RG 1 ∑ ( k+1 k+1 )2 ( k+1 )2
)
L≡ u̇i − u̇i + v̇k+1 − v̇i (17)
NF A 2 i i

where xki and yki are the horizontal and vertical position of node i at time
step k; rijk is the distance between nodes i and j at time step k. In order to where NF represents the number of fluid nodes.
prevent error accumulation in long-term predictions, random-walk
noise was added to all node positions. 2.4. GNS with pressure estimation (GNS-P)
Then, the features are passed over the three functions of GNS:
encoder, processor, and decoder. Each function has a fully connected This study also presented GNS-P, which calculated both particle
neural network (NN). First, the encoder generates a latent graph from motion and pressure. In GNS-P, the node feature was enriched as
the node and edge features, [ ]
nki ≡ u ̃ki , u
̃k−i 1 , ⋯u ̃k , ̃
̃k−i 4 , Δ k
̃k− 1 , ⋯̃k− 4
(18)
〈 k 〉0 ( ) wall,i pi , pi pi
ni = fenc(n) nki (10)

5
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 6. Time histories of the nondimensional pressure at the measurement points with height YP = 0.003, 0.015, 0.030, 0.080 [m] on the side wall. The experiment
data were collected by Lobovský et al. (2014). The pressure estimated with the latest GNN-based particle method (ISPH_GNN; Zhang et al., 2024a) is also plotted as
a reference.

Fig. 7. MSE of the pressure calculated with MPS or GNS-WP against the reference pressure measured by Lobovský et al. (2014). MSE is divided into the square of the
mean error μ2e and the variance σ2e .

k pki (14), the predicted vector was defined as

where ̃
pi ≡
P [ k+1 k+1 k+1 ]
u̇i v̇ p
zk+1 ≡ , i , i (19)
where pki is the pressure of node i at time step k. The current and past four
i
A A P
pressures are given after the normalization with the representative
Unlike common particle methods, we did not detect free surface nodes
pressure P. Random-walk noise was added to the pressures in the same
or impose Dirichlet boundary condition onto the free surface in GNS-P
manner as the positions when computing the node feature. Instead of Eq.

6
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Table 2 2.5. GNS with wall boundary nodes and pressure estimation (GNS-WP)
Time differences between the occurrence of the pressure peaks. The timings of
the peak pressure calculated with MPS or GNS-WP are compared with those in This study finally introduced GNS-WP, the combination of GNS-W
the experiment (Lobovský et al., 2014). A positive value indicates the calculated and GNS-P. The node feature was updated as follows:
peak occurred later than the experiment, while a negative value means the
[ k k− 1 k− 4 ]
opposite. The three cases, YP = 0.080 [m] in scenario A2 and YP = 0.030, nki ≡ ũi , u ̃k−i 4 , qi , ̃
̃i , ⋯u k
pi , ̃
k− 1
pi , ⋯̃
pi (21)
0.080 [m] in scenario A3, have been excluded because they involved a gradual
pressure increase and no distinct peak. The predicted vector was the same as Eq. (19). The loss function can be
Scenario YP [m] MPS − Exp. [s] GNS-WP − Exp. [s] computed as follows:
A2 0.003 0.046 − 0.199 NF ( N
c ∑ ( k+1 k+1 )2 ( k+1 )2
) 1 ∑ ( k+1 )2
​ 0.015 0.232 − 0.221 L≡ u̇i − u̇i + v̇k+1 − v̇i + 2 pi − pk+1
​ 0.030 0.302 − 0.001 NF A i
2 i
NP i i

0.080 N/A N/A

​
(22)
A3 0.003 0.297 0.501
​ 0.015 − 0.792 − 0.266
​ 0.030 N/A N/A
​ 0.080 N/A N/A 2.6. Dataset generation and training

With the MPS solver stated in Section 2.1, we generated four kinds of
(and GNS-WP). It was expected that the boundary condition would be datasets: dam-break, sloshing, hydrostatic, and free-oscillation-in-a-
implicitly considered through learning from the MPS data. The pressure triangular-tank datasets, which are referred to as Datasets A, B, C, and
loss was added to the acceleration loss as follows: D, respectively. Fig. 3 depicts the schematics of the problem settings. In
N ( N the benchmark experiments by Lobovský et al. (2014), Hu and Kashi-
c ∑ ( k+1 k+1 )2 ( k+1 )2
) 1 ∑ ( k+1 )2
L≡ u̇i − u̇i + v̇k+1 − v̇i + 2 pi − pk+1 wagi (2004), and Kishev et al. (2006), pressure was measured at the
NA2 i i
NP i i
point(s) with height YP on the side wall. In the experiment by Martin and
(20) Moyce (1952), the horizontal position of the water front X and the
height of the residual water column Y were studied. On the hydrostatic
where pk+1
i is the ground-truth pressure of particle i at time step k +1 test, the pressure at the center of the bottom wall was compared with the
obtained by particle methods; c represents a tuning parameter for theoretical solution. On the sloshing test (Kishev et al., 2006), the
balancing the acceleration and pressure losses. rectangular tank was forced to oscillate in the sway direction with an
amplitude Ξ and a period T. On the free-oscillation-in-a-triangular-tank
test (Lamb, 1932), from an initial state with the water surface tilted by θ,

Fig. 8. Comparison of the snapshots at t = 1.300, 1.400, 1.725, 1.900 [s]. In MPS, GNS-P, and GNS-WP, particles or nodes are colored by their pressure.

7
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 9. Time histories of the pressure at the measurement point with height YP = 0.100 [m] on the side wall. The experiment data were collected by Kishev
et al. (2006).

the surface profile was compared with the analytical solution. Table 1 6
(Kingma and Ba, 2014) with the learning rate α = 0.1l/(5×10 )+4 , where l
lists the condition of each dataset. For training in the dam-break cases, is the epoch count. Mini batch learning with a batch size of 2 was
nine datasets (Datasets A4‒12) were used, whereas for training in the conducted.
sloshing, one dataset (Dataset B1) was used. Note that the previous
studies (e.g. Zhang et al., 2023 and 2024a) commonly prepared suffi- 3. Results
cient training data that resembled the target problem. However, this is
not efficient; a more desirable approach would solve various problems 3.1. Dam-Break simulations
using minimal training data. To verify the versatility of the proposed
methods when given limited training data, we prepared only one dataset The proposed methods, GNS, GNS-W, GNS-P, and GNS-WP, were first
for sloshing. Unless stated otherwise, the MPS solver employed the verified on dam-break tests. The models were trained using Datasets
[ ] [ ]
following parameters: ρ = 1000 kg /m3 , ν = 1 × 10− 6 m2 /s , g = A4‒12 and tested using Datasets A1‒3. The node and time step sizes
[ ]
9.81 m /s2 , the fixed time step size Δt = 1 × 10− 4 [s], particle size d = were the same as those in training. Fig. 4 shows the time histories of the
5 × 10− 3 [m], and effective radius R = 2.4 d [m]. The Courant number nondimensional horizontal position of the water front and height of the
was set as CCFL = 0.2 [ − ], which is used in RF-SDS (Iida and Yokoyama, residual water column in the test case of A1. The numerical results are
2022b). Each dam-break dataset contained one-second data with 10,000 compared with the experiment data (Martin and Moyce, 1952) and the
frames, while the sloshing dataset contained 11-seconds data with 110, existing surrogate model named ISPH_ML (Zhang et al., 2023). The re-
000 frames. sults by GNS-P and GNS-WP showed good agreement with those by MPS.
Using these datasets, GNS, GNS-W, GNS-P, and GNS-WP were inde- This indicates that the pressure estimation played an important role in
pendently trained. Although, in GNN, a node does not have the “size,” in predicting the fluid motion. It was also confirmed that GNS-P and
this study, we refer to the initial node spacing as “node size” for con- GNS-WP reproduced the experiment data with accuracy comparable to
venience. Unless otherwise mentioned, the node size was the same as the ISPH_ML.
particle size of MPS: dG = d = 5 × 10− 3 [m]. We used the time step size Fig. 5 shows the comparisons of the snapshots at t = 0.24, 0.36,
ΔtG = 1 × 10− 3 [s], which was ten times that of the training data. The 0.48, 0.60 [s] in scenario A2 and t = 0.20, 0.40, 0.60, 0.80 [s] in sce-
convergence study on the time step and node sizes will be presented in nario A3. Overall, it can be observed that all methods roughly repro-
Section 3.3 and Section 3.4. The connectivity radius was fixed as RG = duced the MPS result. Comparing the results of MPS and GNS, there are
2.4dG [m]. The representative values for normalization, U, A, and P, differences in the timing of the fluid reaching the right wall and the
were determined as the standard deviations of the velocities, accelera- height of the splashes. This issue was resolved by introducing wall
tions, and pressures in the training data: for the dam break, U = boundary nodes, but the fluid computed by GNS-W shows slightly un-
[ ]
0.639 [m /s], A = 19.2 m /s2 , and P = 1175 [Pa]; for the sloshing, U = natural behavior, such as the bulge around x = 0.4 [m]. On the other
[ 2
] hand, GNS-P and GNS-WP reproduced the fluid behavior well, con-
0.299 [m /s], A = 12.7 m /s , and P = 499 [Pa]. For stable computa-
firming that the pressure estimation contributed to accurate fluid
tion, the random-walk noises following normal distributions
( ) reproduction. Further, the disturbance in pressure seen in GNS-P was
N 0, 2.5 ×A ×ΔtG2 and N(0, P /10) were added to the predicted posi-
suppressed in GNS-WP. This suggests that the wall nodes might help
tion and pressure of each node, respectively. The tuning parameter for stabilize pressure calculations. Comparing the MPS and GNS-WP results,
balancing losses were given as c = 0.1. the water level in GNS-WP seems slightly higher, and the pressure is
Each surrogate model, GNS, GNS-W, GNS-P, and GNS-WP, was somewhat lower, especially at t = 0.360 [s] in scenario A2. This is likely
composed of five NNs: fenc(n) , fenc(e) , fmsg , fupd , and fdec . Each NN had two because GNS-WP did not explicitly impose the incompressibility con-
hidden layers with 128 nodes and ReLU activations. The message dition and thus allowed for slight volumetric changes.
passing repeated M = 6 times. The NNs were optimized using Adam

8
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 10. Comparison of the snapshots generated by the MPS solver and GNS-WP with different time step sizes: ΔtG = 0.0001, 0.0005, 0.0010, 0.0020, 0.0050,
0.0100 [s]. The corresponding Courant numbers were CCFL = 0.09, 0.45, 0.90, 1.80, 4.48, 8.90 [ − ].

Fig. 6 demonstrates the time histories of the nondimensional pres- shows the mean squared error (MSE) of the pressure calculated with
sure at the measurement points with height YP = 0.003, 0.015, 0.030, MPS or GNS-WP against the reference pressure measured by Lobovský
0.080 [m] on the side wall. The pressure calculated by GNS-WP et al. (2014). According to Iida (2024), (root) MSE can be divided into
matched well with that by MPS and showed similar trends with the mean and variance of the errors as follows:
ISPH_GNN (Zhang et al., 2024a), which is the latest surrogate solver N
based on GNN. GNS-WP seemingly suppressed pressure fluctuation, 1 ∑
MSE = (pi − pi )2 = μ2e + σ 2e (23)
which is commonly seen in particle methods. This occurred because MPS N i
solves PPE for all particles and thus likely involves momentary pressure
N N
spikes due to particle proximity, whereas GNS-WP minimizes the 1 ∑ 1 ∑
where μe = (pi − pi ), σ 2e = (pi − pi − μe )2
accumulative error against the reference pressure and thus tends to N i N i
reduce sensitivity to particle arrangement. In addition, the random-walk
noises to the positions and pressures were considered to prevent over- Here, pi represents the reference pressure. To investigate errors deeply,
fitting to the training data and smooth the estimated pressure. Fig. 7 it was recommended to show not only the value of MSE/RMSE but also

9
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 11. Time histories of the pressure at the measurement point with height YP = 0.100 [m] on the side wall in the test case B2. The pressure calculated by MPS with
Δt = 0.0001 [s] and GNS-WP with ΔtG = 0.0001, 0.0005, 0.0010, 0.0020, 0.0050, 0.0100 [s] are compared. The experiment data is not displayed for visual clarity.

Table 3
Computation time to perform the sloshing simulation for one second. As the MPS
did not support GPU computation, GNS-WP was calculated with both CPU-only
and GPU-based configurations.
Method Time Step Size Courant Number Time (CPU) Time (GPU)
[s] [-] [s] [s]

MPS 0.0001 0.09 2727.92 N/A

GNS- 0.0001 0.09 3803.22 105.36
WP
​ 0.0005 0.45 805.73 21.06
​ 0.0010 0.90 433.77 11.19
​ 0.0020 1.80 193.23 5.60
​ 0.0050 4.48 77.40 2.32
​ 0.0100 8.90 43.00 1.05

the mean and variance components separately. Comparing the errors by

MPS and GNS-WP, in total, GNS-WP shows smaller MSE than MPS,
except for the case of YP = 0.003 [m] in scenario A3. This exception is
Fig. 12. MSE of the pressure calculated with MPS or GNS-WP against the likely due to the time differences between the occurrence of the pressure
reference pressure measured by Kishev et al. (2006). MSE is divided into the peak; in this case, GNS-WP mistakenly estimated the initial peak pres-
square of the mean error μ2e and the variance σ2e . sure values as zero, which greatly affected the MSE. Looking at the
breakdown of MSE, the mean components μ2e are similar between MPS
and GNS-WP, whereas the variance components σ 2e tend to differ. While
both MPS and GNS-WP captured the trends of the reference pressure,
GNS-WP had less error variation. This can be attributed to its ability to
suppress the pressure fluctuation, as previously mentioned. Table 2
shows the time differences between the occurrence of the pressure
peaks, where the timings of the peak pressure calculated with MPS or
GNS-WP are compared with those in the experiment (Lobovský et al.,
2014). In the scenario with a higher initial water level (A2), GNS-WP
showed the earlier peak than both MPS and the experiment, whereas in
the scenario with a lower initial water level (A3), it showed later. A
possible cause could be slight volumetric changes, but further investi-
gation is required.

3.2. Sloshing simulations

Next, we verified GNS, GNS-W, GNS-P, and GNS-WP on a sloshing

Fig. 13. MSE of the pressure calculated by GNS-WP against that by MPS. MSE
test, involving three major challenges: a moving wall boundary, long-
is divided into the square of the mean error μ2e and the variance σ2e . term prediction, and limited training data. The models were trained
with Dataset B1 and tested with Dataset B2. The node and time step sizes
were the same as those in training. Fig. 8 compares the snapshots at t =

10
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 14. Time histories of the pressure at the measurement point with height YP = 0.100 [m] on the side wall in the test case B2. The pressure calculated by MPS with
d = 5 [mm] and GNS-WP with dG = 5, 6, 7.5, 10 [mm] are compared. The experiment data is not displayed for visual clarity.

1.300, 1.400, 1.725, 1.900 [s]. It can be seen that GNS and GNS-P were
unable to capture the moving wall boundary at all. Although GNS-W
resolved this issue, it failed to simulate the shape of free surface well,
especially at the transient phase around t = 1.725 [s]. In contrast, GNS-
WP reproduced the violent sloshing flow well, including the free-surface
profiles and pressure distributions. This indicates that wall nodes are
required to treat moving wall boundaries, and that accurate pressure
estimation helps in replicating fluids.
Fig. 9 shows the time histories of the pressure at the measurement
point with height YP = 0.100 [m] on the side wall. The pressure pre-
dicted by GNS-WP was in good agreement with both the MPS and
experiment data. GNS-WP managed to simulate the sloshing problem,
which involved violent free-surface flow and a moving wall boundary
for over a long time (11 s, totaling 11,000 predictions), when given
limited training data. The pressure estimation will be evaluated quan-
titatively in Section 3.3.

3.3. Convergence study on time step size

Fig. 15. MSE of the pressure calculated with MPS or GNS-WP against the
reference pressure measured by Kishev et al. (2006). MSE is divided into the In this section, we examine the convergence property of GNS-WP
square of the mean error μ2e and the variance σ2e . with respect to time step sizes. The models of GNS-WP were trained
with Dataset B1 and tested with B2. To investigate the effect of Courant
number, the time step sizes were set as ΔtG = 0.0001, 0.0005, 0.0010,
0.0020, 0.0050, 0.0100 [s], while fixing the node size to dG = 5 ×
10− 3 [m]. The corresponding Courant numbers were CCFL = 0.09, 0.45,
0.90, 1.80, 4.48, 8.90 [ − ]. If the time step size is sufficiently small, the
position noise affects more significantly; therefore, the random-walk
noise was generated according to a normal distribution
( )
N 0, 0.75 ×A ×ΔtG2 in the case of ΔtG = 0.0001 [s]. Fig. 10 compares
the snapshots generated by the MPS solver and GNS-WP with different
time step sizes. Except for the cases with ΔtG = 0.0050 [s] or ΔtG =
0.0100 [s], the results show good agreement with MPS. Fig. 11 shows
the time histories of the pressure at the measurement point, and Fig. 12
demonstrates the MSE of the pressure calculated with MPS or GNS-WP
against the experimental value. When using the same time step size
(ΔtG = Δt = 0.0001 [s]), GNS-WP predicted the pressure with smaller
errors, especially the variance σ 2e , than MPS. This is likely because GNS-
WP suppressed pressure fluctuation more effectively than MPS. For
ΔtG ≤ 0.0020 [s], i.e., CCFL ≤ 1.80 [ − ], the MSE by GNS-WP was com-
Fig. 16. MSE of the pressure calculated by GNS-WP against that by MPS. MSE parable to that by MPS, implying that GNS-WP may allow for signifi-
is divided into the square of the mean error μ2e and the variance σ2e . cantly larger Courant numbers than conventional particle methods.
Fig. 13 shows the MSE of the pressure calculated by GNS-WP against that
by MPS. It has a similar tendency that the errors were suppressed for
ΔtG ≤ 0.0020 [s]. However, it should be noted that the mean component

11
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 17. Comparison of the snapshots at t = 0.2, 0.4, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6 [s] in the dam-break scenario (A13).

of the MSE, μ2e , was relatively large in the case of ΔtG = 0.0020 [s] in i9–13900) and a GPU (NVIDIA(R) GeForce RTX 4080). GNU Fortran
Fig. 12, which suggests that the trends of the reference pressure were not 11.4.0 was used for MPS, whereas PyTorch 2.2.1 on Python 3.8 (and
precisely predicted. CUDA 12.3 for GPU computation) were used for GNS-WP. Table 3 lists
Furthermore, the computation times were compared on the sloshing the computation time to simulate the sloshing for one second. As the
test (scenario B2), where 2,880 fluid particles/nodes and 1,116 wall MPS solver did not support GPU computation, GNS-WP was calculated
particles/nodes were used. This test was conducted on a common PC with both CPU-only and GPU-based configurations for comparison.
(Ubuntu 22.04 LTS) equipped with a CPU (13th Gen Intel(R) Core Notably, GNS-WP significantly reduced computation time by increasing

12
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 18. Time histories of the pressure at the measurement point with height YP = 0.010 [m] on the side wall. The experiment data were collected by Hu and
Kashiwagi (2004).

3.4. Convergence study on node size

This section presents the convergence study on GNS-WP with regard

to node sizes. We used the same model as that described in Section 3.3,
with the time step size ΔtG = 0.0001 [s] and the node size dG =
0.005 [m]. The model was tested in the sloshing scenario (B2) using
different node sizes: dG = 0.005, 0.006, 0.0075, 0.010 [m]. Figs. 14‒16
show the time histories of the pressure at the measurement point, the
MSE of the pressure calculated with MPS or GNS-WP against the
experimental value, and the MSE of the pressure calculated by GNS-WP
against that by MPS, respectively. These figures suggest that as the node
size decreased, the errors against both the experiment and MPS were
reduced. A certain level of convergence can be seen with respect to the
node size, and the accuracy was maintained even with the node size
increased to dG = 0.006 [m], or 1.2 times the original size. However, at
present, it appears challenging to substantially enlarge the node size,
unlike the time step size.

Fig. 19. MSE of the pressure calculated with MPS or GNS-WP against the
3.5. Validation of applicability
reference pressure measured by Hu and Kashiwagi (2004). MSE is divided into
the square of the mean error μ2e and the variance σ2e .
This section investigates the generalization performance of GNS-WP.
The model trained on the sloshing (scenario B1) was applied to three
the time step size and using a GPU. By selecting ΔtG = 0.001 [s], i.e.,
different problems: the dam-break (scenario A13), hydrostatic (sce-
CCFL = 0.90 [ − ], the proposed method can perform one-second pre-
narios C1‒2), and free-oscillation-in-a-triangular-tank (D1) tests. Note
diction with sufficient accuracy in just 11 s. When using the same time
that we consistently used the same model as that described in Section
step size as MPS, GNS-WP took longer CPU time than MPS. This could be
3.3, with the time step size ΔtG = 0.001 [s] and the node size dG =
due to the difference in processing speeds between Fortran and Python.
0.005 [m].
As shown in Appendix A, the computation time of GNS-WP tends to
Fig. 17 demonstrates the comparison of the snapshots between MPS
increase linearly with the number of nodes, and thus it should scale more
and GNS-WP in the dam-break test (A13). GNS-WP accurately simulated
effectively as the number of particles increases.
the timing of reaching the side wall, shape of the breaking wave, and

13
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

while in GNS-WP, it was 0.011 [s] later. Fig. 19 compares the MSE of the
pressure calculated with MPS and GNS-WP against the reference. These
figures suggest that GNS-WP predicted hydrodynamic pressure with
accuracy equivalent to MPS while mitigating the pressure fluctuation. As
a result, the proposed method successfully reduced the error relative to
the experiment, even without learning from the dam-breaking data.
Notably, the CPU computation time for MPS was 5,032 [s] per second,
whereas GNS-WP required 510 [s]. With GPU acceleration, this was
further reduced to 24 [s], achieving more than 200 times the speed than
MPS.
Next, the same GNS-WP model was applied to hydrostatic cases.
Figs. 20 and 21 show the snapshots in scenarios C1 and C2 and the time
histories of the nondimensional pressure at the center of the bottom
wall, respectively. MPS showed good agreement with the theoretical
pressure, though it involved gradual decrease because of imperfect
volume conservation, as reported by Iida and Yokoyama (2022b).
GNS-WP had similar trends with MPS in the case of YF = 0.1 [m],
whereas it differed in the case of YF = 0.2 [m]. As the only difference
between the two cases was the initial water depth, the GNS-WP model
might be unable to estimate pressure beyond a certain level. In fact, only
0.25 % of the fluid particles in the training data had pressures greater
than p = 1962 [Pa], indicating that the model may not have sufficiently
learned to handle such high pressure. In summary, at this stage, GNS-WP
has difficulty in solving problems with unknown pressure levels.
Finally, we tested the GNS-WP model in scenario D1, where the fluid
freely oscillates across a channel whose section consists of two straight
lines inclined at 45∘ to the vertical. According to Lamb (1932), the free
surface always forms a straight line in the slowest mode of oscillation.
On this test, the initial water depth at the center was set as YF = 0.14 [m]
to keep the pressure distribution within the range of the training data,
and the particle/node sizes were set to d = dG = 0.004 [m] to retain
sufficient spatial resolution. With an initial tilt angle of θ = tan− 1 (0.2),
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
the maximum water depth is given by h = YF / 1 − tan2 θ ≈ 0.1429 [m]
at the moment the surface becomes horizontal. Fig. 22 depicts the nu-
merical results with the theoretical free surface profiles at t = T/8, T/4,
T/2, T, where T is the natural period. Note that, as GNS-WP was trained
on viscous fluid data, it does not strictly match the analytical solutions.
Therefore, the results obtained from both inviscid and viscous MPS are
Fig. 20. Comparison of the snapshots in the test cases of (a) C1 and (b) C2, of
presented for reference. The inviscid MPS reproduced straight free sur-
which difference is only the initial water depth: YF = 0.1, 0.2 [m].
faces consistent with the analytical solutions, despite showing slight
volumetric changes. GNS-WP generated surface profiles and pressure
pressure distribution. Fig. 18 shows the time histories of the pressure at
distributions similar to the viscous MPS, indicating its applicability to a
the measurement point with height YP = 0.010 [m] on the side wall. The
wall boundary different from that in the training data. On the other
peak pressure in MPS occurred 0.009 [s] earlier than in the experiment,
hand, the fluid volume seems to decrease compared to MPS. This point

Fig. 21. Time histories of the nondimensional pressure at the measurement point (the center of the bottom wall. The pressure is normalized with theoretical values: p
= 981 [Pa] in scenario C1 and p = 1962 [Pa] in scenario C2.

14
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 22. Comparison of the snapshots at t = 0.095, 0.189, 0.379, 0.758 [s], i.e., t = T/8, T/4, T/2, T. The gray dashed lines represent the theoretical free surface
profiles for inviscid fluids. As GNS-WP was trained on viscous fluid data, it does not strictly match the analytical solutions. Therefore, the results obtained from both
inviscid and viscous MPS are presented for reference.

Fig. 23. Comparison of the snapshots generated by the MPS solver and GNS-WP in the test case A3. The particles or nodes are colored by their velocity divergence.

will be discussed in Section 4. and 26 depict the distributions of the normalized PND, ni /n0 , where the
PND of particle i is divided by the representative PND, n0 . While the
4. Discussions PNDs were kept constant in MPS, except near the boundaries, GNS-WP
showed some variation, indicating expansion and contraction of fluids.
In the prior sections, our proposed surrogate models tended to This is because the incompressibility condition was not explicitly
involve the changes in fluid volume. Then, we investigated the physical imposed in the GNS-WP algorithm.
consistency of GNS-WP, such as the conservation of mass/volume and Fig. 27 shows the time histories of the normalized mechanical energy
energy. First, regarding the mass/volume conservation, we checked the of all fluid particles/nodes in the test case A3. The potential, kinetic, and
incompressibility from two perspectives: the divergence-free and parti- mechanical energy was divided by E0 , the total energy at the initial
cle number density (PND) conditions. Figs. 23 and 24 show the velocity- condition. Note that MPS does not strictly conserve energy because it has
divergence distributions in the test cases A3 and B2, respectively. In the a viscosity term. Although the energy trends in MPS and GNS-WP were
dam-break test (A3), both MPS and GNS-WP showed similar trends: the generally consistent, GNS-WP did not satisfy the energy conservation at
initially zero velocity divergence was disturbed near the surge front, this stage; for example, as the water column began to collapse, the po-
surrounded by the free-surface and wall boundaries. In the sloshing test tential energy surpassed that at the initial state. This is probably due to
(B2), GNS-WP showed similar tendencies to those in MPS unless ΔtG was the slight expansion of the fluid caused by compressibility, leading to an
too large. In both cases, the disturbances in GNS-WP appear to be increase in volume. This issue should also be addressed in future
mitigated compared to MPS. This is likely for the same reason that research.
pressure fluctuations were suppressed in the GNS simulations. Figs. 25

15
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 24. Comparison of the snapshots generated by the MPS solver and GNS-WP with different time step sizes in the test case B2. The particles or nodes are colored by
their velocity divergence.

16
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 25. Comparison of the snapshots generated by the MPS solver and GNS-WP in the test case A3. The particles or nodes are colored by their normalized PND.

5. Conclusions methodology and results. The detailed investigation into the feature
settings will be beneficial for the further development of surrogate
This paper proposed particle-based surrogate models applicable to models. Moreover, this study is the first to demonstrate that the time
larger time step sizes and different fluid phenomena. In this study, we step size of DNN-based Lagrangian solvers can be significantly increased
introduced three GNS-based surrogate models: GNS-W, which incorpo- while fixing the particle/node size. The convergence study showed that
rated wall boundary nodes; GNS-P, which enabled pressure estimation; GNS-WP may allow for larger Courant numbers than conventional
and GNS-WP, a combination of both. Through the step-by-step verifi- methods, suggesting the potential for drastically reducing computation
cation with these models, we revealed which features are important for time. Further, no prior studies have applied a surrogate model trained on
accurate simulations of Lagrangian fluids. First, the four surrogate a fluid phenomenon to different ones, such as from sloshing to dam
models, GNS, GNS-W, GNS-P, and GNS-WP, were verified on the dam- break, gentle free oscillation, and hydrostatic problems. This is a unique
break tests. We demonstrated that the pressure estimation contributed finding of our research.
to accurate fluid reproduction, and the wall nodes might help stabilize In future work related to this study, we plan to address the following
pressure calculations. Then, we verified the four models on the sloshing four challenges:
test, involving three major challenges: a moving wall boundary, long-
term prediction, and limited training data. The comparative study 1) Insufficient physical consistency: the proposed surrogate model is
indicated that accurate pressure estimation helps in replicating fluids, not guaranteed to satisfy physical laws, such as the conservation of
and that wall nodes are required to treat moving wall boundaries. Even mass, momentum, and energy, and the divergence-free condition.
though the training data was limited to only one case, GNS-WP managed This is likely because important indicators such as PND and velocity
to simulate the sloshing problem over 11 s, totaling 11,000 predictions. divergence were not included as features. In addition, to satisfy
Additionally, we examined the convergence property of GNS-WP with momentum conservation, the reaction forces between two nodes
respect to time step sizes, while fixing the node size. Notably, the pro- must be antisymmetric, but the proposed method does not currently
posed method reproduced the hydrodynamic pressure on the sloshing ensure this condition. Future work will focus on developing physi-
test with sufficient accuracy, using the time step ten times as large as cally consistent surrogate models.
that of the training data (MPS). Furthermore, we investigated the 2) Management of unknown pressure: the current surrogate model
model’s convergence property regarding the node size. A certain level of tends to struggle with solving problems that involve pressure levels
convergence can be seen, and the accuracy was maintained even with higher than those in the training data. By incorporating gradients
the node size increased 1.2 times the original. Finally, the model of GNS- and/or time derivatives of pressure into the learning process, the
WP, trained in the sloshing scenario, was validated on three different model is expected to treat unknown pressure and enhance its
problems: the dam-break, hydrostatic, and free-oscillation-in-a- robustness and generalization performance.
triangular-tank tests. In the dam-break simulation, the proposed 3) Extension to complex and/or three-dimensional problems: at pre-
method achieved accuracy equal or better than MPS, with ten times the sent, this study is limited to validation on the basic benchmark tests
speed on CPU and over 200 times faster on GPU. Even though only a in 2D. For practical engineering applications, the applicability to
violent sloshing flow was used during training, GNS-WP successfully more complex problems, such as fluid-structure interaction or
reproduced both static and calm flows with a different wall boundary. slamming problems in 3D.
One of the major contributions of this study is that it presented the 4) Establishment of experimental-data-driven surrogate models: the
step-by-step improvement of the particle-based surrogate models method proposed in this study has the potential to learn from
although most of the previous studies have shown only the final experimental data measured in the real-world. This approach could

17
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 26. Comparison of the snapshots generated by the MPS solver and GNS-WP with different time step sizes in the test case B2. The particles or nodes are colored by
their normalized PND.

18
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. 27. Time histories of the normalized mechanical energy in the test case A3. Each energy component was divided by the total energy at the initial condition E0 .

reproduce complicated fluid phenomena of which the governing interests or personal relationships that could have appeared to influence
equations are not known, such as multiphase flows involving gran- the work reported in this paper.
ular and discontinuous materials. This study is considered an
important foundation for such future studies. Acknowledgement

CRediT authorship contribution statement This work was supported by JSPS KAKENHI (Grant Number
24K22934).
Takefumi Higaki: Writing – original draft, Visualization, Supervi-
sion, Software, Project administration, Methodology, Investigation, Declaration of generative AI and AI-assisted technologies in the writing
Formal analysis, Data curation, Conceptualization. Yuki Tanabe: process
Writing – review & editing, Visualization, Validation, Software, Meth-
odology, Investigation, Formal analysis, Data curation. Hirotada During the preparation of this work the authors used ChatGPT (GPT-
Hashimoto: Writing – review & editing, Supervision, Resources, 4o) in order to improve language and readability. After using this tool/
Methodology, Conceptualization. Takahito Iida: Writing – review & service, the authors reviewed and edited the content as needed and take
editing, Software, Investigation. full responsibility for the content of the publication.

Declaration of competing interest

The authors declare that they have no known competing financial

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.apor.2025.104424.

Appendix A. Scaling Benefit of GNS-WP

This section shows a brief study of the computation time with respect to the number of particles/nodes. Fig. A.1 shows the computation time
measured on the dam-break tests (scenarios A1, A2, and A12) with the particle size d = dG = 5 [mm] and the time step size Δt = ΔtG = 1 × 10− 3 [s].
Koshizuka et al. (2018) reported that the computation time for solving PPE, the most time-consuming process in MPS, is proportional to the 1.5th
power of the number of particles when using iterative methods like incomplete Cholesky conjugate gradient method. On the other hand, the proposed
surrogate model does not solve PPE, and its computation time increased linearly with the number of nodes. Thus, the surrogate model is expected to
offer significant scaling benefits as the number of particles increases.

19
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Fig. A.1. Computation time to simulate the dam break for one second by MPS (CPU), GNS-WP (CPU) and GNS-WP (GPU) with different number of particles/nodes.

References Koshizuka, S., Oka, Y., 1996. Moving-particle semi-implicit method for fragmentation of
incompressible fluid. Nucl. Sci. Eng. https://doi.org/10.13182/NSE96-A24205.
Koshizuka, S., Shibata, K., Kondo, M., Matsunaga, T., 2018. Moving Particle Semi-
Bai, J., Zhou, Y., Ma, Y., Jeong, H., Zhan, H., Rathnayaka, C., Sauret, E., Gu, Y., 2022.
Implicit method: a Meshfree Particle Method For Fluid Dynamics. Academic Press.
A general Neural Particle Method for hydrodynamics modeling. Comput. Methods
Ladický, L., Jeong, S., Solenthaler, B., Pollefeys, M., Gross, M., 2015. Data-driven fluid
Appl. Mech. Eng. 393, 114740. https://doi.org/10.1016/j.cma.2022.114740.
simulations using regression forests. ACM Trans. Graph. 34, 1–9. https://doi.org/
Chorin, A.J., 1968. Numerical solution of the Navier-stokes equations. Math. Comput.
10.1145/2816795.2818129.
22, 745. https://doi.org/10.2307/2004575.
Lamb, H., 1932. Hydrodynamics, 6th edition. Cambridge University Press, pp. 442–443.
Cummins, S.J., Rudman, M., 1999. An SPH projection method. J. Comput. Phys. 152,
Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A., 2018. Learning particle
584–607. https://doi.org/10.1006/jcph.1999.6246.
dynamics for manipulating rigid bodies, deformable objects, and fluids. In:
Duan, G., Chen, B., Zhang, X., Wang, Y., 2017. A multiphase MPS solver for modeling
Proceedings of the International Conference on Learning Representations (ICLR
multi-fluid interaction with free surface and its application in oil spill. Comput.
2019).
Methods Appl. Mech. Eng. 320, 133–161. https://doi.org/10.1016/j.
Li, Z., Farimani, A.B., 2022. Graph neural network-accelerated Lagrangian fluid
cma.2017.03.014.
simulation. Comput. Graph. 103, 201–211. https://doi.org/10.1016/j.
Gingold, R.A., Monaghan, J.J., 1977. Smoothed particle hydrodynamics: theory and
cag.2022.02.004.
application to non-spherical stars. Mon. Not. R. Astron. Soc. 181, 375–389. https://
Lo, E.Y.M., Shao, S., 2002. Simulation of near-shore solitary wave mechanics by an
doi.org/10.1093/mnras/181.3.375.
incompressible SPH method. Appl. Ocean Res. 24, 275–286. https://doi.org/
Hu, C., Kashiwagi, M., 2004. A CIP-based method for numerical simulations of violent
10.1016/s0141-1187(03)00002-6.
free-surface flows. J. Mar. Sci. Technol. 9, 143–157. https://doi.org/10.1007/s00
Lobovský, L., Botia-Vera, E., Castellana, F., Mas-Soler, J., Souto-Iglesias, A., 2014.
773-004-0180-z.
Experimental investigation of dynamic pressure loads during dam break. J. Fluids
Iida, T., 2024. Identifying causes of errors between two wave-related data using
Struct. 48, 407–434. https://doi.org/10.1016/j.jfluidstructs.2014.03.009.
performance metrics. Appl. Ocean Res. 148, 104024. https://doi.org/10.1016/j.apo
Luo, M., Khayyer, A., Lin, P., 2021. Particle methods in ocean and coastal engineering.
r.2024.104024.
Appl. Ocean Res. 114, 102734. https://doi.org/10.1016/j.apor.2021.102734.
Iida, T., Yokoyama, Y., 2022a. Investigation of numerical conditions of moving particle
Marrone, S., Colagrossi, A., Le Touzé, D., Graziani, G., 2010. Fast free-surface detection
semi-implicit for two-dimensional wedge slamming. J. Mar. Sci. Appl. 20, 585–594.
and level-set function definition in SPH solvers. J. Comput. Phys. 229, 3652–3663.
https://doi.org/10.1007/s11804-021-00234-x.
https://doi.org/10.1016/j.jcp.2010.01.019.
Iida, T., Yokoyama, Y., 2022b. Convergence-improved source term of pressure Poisson
Martin, J.C., Moyce, W.J., 1952. Part IV. An experimental study of the collapse of liquid
equation for moving particle semi-implicit. Appl. Ocean Res. 124, 103189. https://
columns on a rigid horizontal plane. Philos. Trans. R. Soc. Lond. 244, 312–324.
doi.org/10.1016/j.apor.2022.103189.
https://doi.org/10.1098/rsta.1952.0006.
Iribe, T., Nakaza, E., 2011. An improvement of accuracy of the mps method with a new
Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L., Tenenbaum, J.B., Yamins, D.L.
gradient calculation model. J. Jpn. Soc. Civ. Eng. Ser. B2 (Coast. Eng.) 67, 36–48.
K., 2018. Flexible neural representation for physics prediction. In: Proceedings of the
https://doi.org/10.2208/kaigan.67.36.
32nd Conference on Neural Information Processing Systems (NeurIPS 2018).
Khayyer, A., Gotoh, H., 2010. A higher order Laplacian model for enhancement and
Prantl, L., Ummenhofer, B., Koltun, V., Thuerey, N., 2022. Guaranteed conservation of
stabilization of pressure calculation by the MPS method. Appl. Ocean Res. 32,
momentum for learning particle-based fluid dynamics. In: Proceedings of the 36th
124–131. https://doi.org/10.1016/j.apor.2010.01.001.
International Conference on Neural Information Processing Systems (NeurIPS 2022).
Khayyer, A., Gotoh, H., 2011. Enhancement of stability and accuracy of the moving
Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., Battaglia, P.W., 2020.
particle semi-implicit method. J. Comput. Phys. 230, 3093–3118. https://doi.org/
Learning to simulate complex physics with graph networks. In: Proceedings of the
10.1016/j.jcp.2011.01.009.
37th International Conference on Machine Learning (ICML 2020).
Kingma, D.P., Ba, J., 2014. Adam: a method for stochastic optimization. In: Proceedings
Schenck, C., Fox, D., 2018. SPNets: differentiable fluid dynamics for deep neural
of the 3rd International Conference on Learning Representations, ICLR 2015.
networks. In: Proceedings of the Conference on Robot Learning, pp. 317–335.
Kishev, Z.R., Hu, C., Kashiwagi, M., 2006. Numerical simulation of violent sloshing by a
Tsuruta, N., Khayyer, A., Gotoh, H., 2013. A short note on dynamic stabilization of
CIP-based method. J. Mar. Sci. Technol. 11, 111–122. https://doi.org/10.1007/
moving particle semi-implicit method. Comput. Fluids 82, 158–164. https://doi.org/
s00773-006-0216-7.
10.1016/j.compfluid.2013.05.001.

20
T. Higaki et al. Applied Ocean Research 154 (2025) 104424

Tsuruta, N., Khayyer, A., Gotoh, H., 2015. Space potential particles to enhance the Zhang, Y., Ban, X., Du, F., Di, W., 2020. FluidsNet: end-to-end learning for Lagrangian
stability of projection-based particle methods. Int. J. Comput. Fluid Dyn. 29, fluid simulation. Expert Syst. Appl. 152, 113410. https://doi.org/10.1016/j.
100–119. https://doi.org/10.1080/10618562.2015.1006130. eswa.2020.113410.
Ummenhofer, B., Prantl, L., Thuerey, N., Koltun, V., 2019. Lagrangian fluid simulation Zhang, N., Yan, S., Ma, Q., Guo, X., Xie, Z., Zheng, X., 2023. A CNN-supported
with continuous convolutions. In: Proceedings of the International Conference on Lagrangian ISPH model for free surface flow. Appl. Ocean Res. 136, 103587. https://
Learning Representations (ICLR 2020). doi.org/10.1016/j.apor.2023.103587.
Wendland, H., 1995. Piecewise polynomial, positive definite and compactly supported Zhang, N., Yan, S., Ma, Q., Li, Q., 2024a. A hybrid method combining ISPH with graph
radial functions of minimal degree. Adv. Comput. Math. 4, 389–396. https://doi. neural network for simulating free-surface flows. Comput. Phys. Commun. 301,
org/10.1007/bf02123482. 109220. https://doi.org/10.1016/j.cpc.2024.109220.
Yao, Q., Wang, Z., Zhang, Y., Li, Z., Jiang, J., 2023. Towards real-time fluid dynamics Zhang, N., Yan, S., Ma, Q., Li, Q., 2024b. Numerical simulation of wave-floater
simulation: a data-driven NN-MPS method and its implementation. Math. Comput. interactions using ISPH_GNN trained on data for wave-only cases. Ocean Eng. 306,
Model. Dyn. Syst. 29, 95–115. https://doi.org/10.1080/13873954.2023.2184835. 118041. https://doi.org/10.1016/j.oceaneng.2024.118041.

21