Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 368 (2020) 113127

www.elsevier.com/locate/cma

The neural particle method – An updated Lagrangian physics

informed neural network for computational fluid dynamics
Henning Wesselsa ,∗, Christian Weißenfelsa,b , Peter Wriggersa
a Institute of Continuum Mechanics, Leibniz Universität Hannover, An der Universität 1, 30823 Garbsen, Germany
b Institute of Applied Mechanics, Technische Universität Braunschweig, Pockelsstraße 3, 38106 Braunschweig, Germany
Received 13 March 2020; received in revised form 8 May 2020; accepted 8 May 2020
Available online 3 June 2020

Abstract
Today numerical simulation is indispensable in industrial design processes. It can replace cost and time intensive experiments
and even reduce the need for prototypes. While products designed with the aid of numerical simulation undergo continuous
improvement, this must also be true for numerical simulation techniques themselves. Up to date, no general purpose numerical
method is available which can accurately resolve a variety of physics ranging from fluid to solid mechanics including
large deformations and free surface flow phenomena. These complex multi-physics problems occur for example in Additive
Manufacturing processes. In this sense, the recent developments in Machine Learning display promise for numerical simulation.
It has recently been shown that instead of solving a system of equations as in standard numerical methods, a neural network
can be trained solely based on initial and boundary conditions. Neural networks are smooth, differentiable functions that can
be used as a global ansatz for Partial Differential Equations (PDEs). While this idea dates back to more than 20 years (Lagaris
et al., 1998), it is only recently that an approach for the solution of time dependent problems has been developed (Raissi
et al., 2019). With the latter, implicit Runge–Kutta schemes with unprecedented high order have been constructed to solve
scalar-valued PDEs. We build on the aforementioned work in order to develop an Updated Lagrangian method for the solution
of incompressible free surface flow subject to the inviscid Euler equations. The method is straightforward to implement and does
not require any specific algorithmic treatment which is usually necessary to accurately resolve the incompressibility constraint.
Due to its meshfree character, we will name it the Neural Particle Method (NPM). It will be demonstrated that the NPM
remains stable and accurate even if the location of discretization points is highly irregular.
⃝c 2020 Elsevier B.V. All rights reserved.

Keywords: Physics-informed neural network; Machine learning; Computational fluid dynamics; Incompressibility; Constraint problem; Implicit
Runge–Kutta

1. Introduction
Today, numerical methods are well established and widely used in research and development. Many different
methods have emerged to tackle a variety of problems. In Computational Fluid Dynamics (CFD), the study of
flows in Eulerian formulation within enclosed domains, e.g. gas turbines, is typically performed with the Finite
Volume (FVM) or the Finite Difference Method (FDM). In these methods, the fluid domain is discretized by a
∗ Corresponding author.
E-mail address: [email protected] (H. Wessels).

https://doi.org/10.1016/j.cma.2020.113127
0045-7825/⃝ c 2020 Elsevier B.V. All rights reserved.
2 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

mesh on which the governing equations are solved. Standard mesh-based methods suffer from three shortcomings
when applied to incompressible fluid flow:
1. In an Eulerian reference frame, convective terms occur and require stabilization [1].
2. To represent continuous free surface flows, mesh-based methods require special techniques such as the
Volume-of-Fluid (VoF) approach [2].
3. The incompressibility constraint causes numerical instabilities and must be stabilized [3].
In order to relax the mesh-dependency and to account for free surface flows, the Arbitrary-Lagrangian–Eulerian
(ALE) formulation has been introduced by Hirt et al. [4], see also [5] and [6]. In the ALE method, the mesh is
advanced in time, but independently of the fluid motion to avoid critical mesh-distortion.
With the aim to entirely remove the mesh-dependency, so-called mesh-free or particle methods have been
developed. These methods usually employ a Lagrangian formulation. The discretization points, referred to as
particles, follow the fluid motion with the effect that convective terms disappear. Using a Lagrangian description,
free surface flows are naturally represented. Besides the well-known Smoothed Particle Hydrodynamics (SPH)
method which was independently introduced by Lucy [7] and Gingold and Monaghan [8], many other schemes
have emerged. Mesh-free methods do not require a fixed mesh, but also in these methods a connectivity is present.
It is established by a search algorithm, which must fulfill certain topological requirements [9]. To equilibrate
unphysical configurational forces, some methods rely on r -adaptivity, e.g. incompressible SPH [10] or the Optimal
Transportation Meshfree Method (OTM) suggested by Li et al. [11]. Alternatively, the equation of motion can be
stabilized, see e.g. [12] for the OTM or [13] for SPH. For an overview of meshfree methods, the reader is referred
to the book of Li and Liu [14].
The incompressible Euler and Navier–Stokes equations solved with either mesh-based or mesh-free methods
have the form of a mixed-method. Mixed methods are subject to the Ladyschenskaja–Babuška–Brezzi (LBB)
or inf–sup condition. It requires the pressure interpolants to be of lower order than those of the velocity [15].
However, in practice equal-order interpolation is often preferred. In this case, the pressure degrees of freedom lay
directly on the fluid boundary. This facilitates the imposition of boundary conditions and the treatment of fluid–
structure interaction. Hence, a variety of stabilization techniques have been developed which circumvent the LBB
condition [15]. According to Brezzi et al. [16], they can be classified according to two different strategies. The first
is to modify the bilinear form in order to achieve enhanced numerical stability without compromising consistency.
The standard Galerkin method is applied to the modified equations. Commonly known methods falling into this
class are the Pressure Stabilized Petrov–Galerkin method (PSPG) proposed by Tezduyar et al. [5] and the Finite
Increment Calculus (FIC) formulation [17]. The second strategy is to enrich the standard Galerkin method with
special functions. More precisely, the space of functions is enlarged such that a comparatively coarse mesh is able
to deal with the effects of unresolvable scales [16]. This approach is followed in the Subgrid Scale Method, in
Residual-free Bubbles or the Variational Multiscale Method, see e.g. [16] and [18].
In addition to the aforementioned numerical methods, data-driven approaches powered by machine learning are
increasingly finding their way into CFD. One field of application is the reconstruction of flow fields from data.
The reconstruction from discrete pressure or velocity measurements can be challenging if the number of available
sensors is limited [19]. Raissi et al. [20] have developed an algorithm which learns the velocity and pressure fields
from continuous flow visualizations such as particle image velocimetry. Data-driven techniques are also applied
on the simulation of fluid flow. For example, neural networks have been developed that accurately resolve the
turbulent Reynolds Averaged Navier Stokes equations. A brief overview can be found in [21]. In the context of
machine learning of turbulent flows, a combination of feature extraction, i.e. Model Order Reduction (MOR) with
Recurrent Neural Networks (RNNs) is commonly employed. MOR techniques to compute the lower dimensional
latent space are for example Proper Order Decomposition (POD) [22] or autoencoder [23]. RNNs are based on
sequence learning and therefore especially suited to learn the latent space dynamics. However, RNNs are difficult
to train in practice. A commonly applied and stable variant of RNNs are Long Short Term Memory networks
(LSTMs) [24]. Another alternative to standard RNNs are Neural Ordinary Differential Equations (NODEs) which
incorporate history implicitly and can learn from temporally scattered data [25]. The performance of NODE and
LSTM in combination with POD applied to Burgers turbulence has been studied in [26]. A detailed overview of
machine learning for fluid dynamics can be found in [27].
A remedy of the aforementioned data-driven simulation concepts is that they are physics agnostic. In order to
increase the fidelity of machine learning for predictive simulations, current research focuses on the implementation
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 3

of physical constraints into formerly purely data-driven approaches. For example, Mohan et al. [28] embed the
incompressibility constraint into a convolutional autoencoder using the equivalence of convolutional neural network
kernels with Finite Volume stencils. While the approach steps in the direction of Physics Informed Data-Driven
Simulation, it has two severe limitations. First, stencils require a fixed lattice. This makes the application of the
method to scenarios like free surface flow and large deformations cumbersome. Second, only the time independent
continuity equation is included as a physical constraint. The solution of the momentum equation remains a
black-box.
Besides their application in data-driven simulation, neural networks can also be used for the numerical solution of
Partial Differential Equation (PDEs) in the absence of data. An innovative approach is to use a feed-forward neural
network as a global ansatz function of a PDE. The universal approximation theorem states that a feed-forward neural
network with mild assumptions on the activation function can approximate any function [29,30]. When a neural
network is designed to take the solution of a PDE as output, it acts as ansatz function of the solution. The idea has
first been presented by Lagaris et al. [31] and can be regarded as a collocation method. Note that a neural network
is a continuous and smooth function which can not only be differentiated with respect to the network parameters
(i.e. the weights and the biases), but also with respect to its input. If the input is the spatial position, derivatives
of the network output with respect to its input are then spatial derivative operators as occurring in the PDE. These
derivatives can be computed with ease even for complex network architectures using Automatic Differentiation
(AD), see e.g. [32] and [33].
While the first attempt of Lagaris et al. [31] was only applicable on rectangular domains, the method has been
extended to more complex geometries [34]. Thanks to the computational power increase since this pioneering work
and the recent developments in machine learning and neural networks, attention in the topic has increased again.
A novel approach to the solution of PDEs on complex domains was presented by Berg and Nyström [35], who
introduced a projection of the actual solution in order to satisfy boundary conditions exactly. For the solution of
quasi-static mechanical problems, Samaniego et al. [36] have used a neural network as ansatz for the displacement
and formulated a loss function that minimizes the elastic energy. The method has been extended from small to
large strain problems by Nguyen-Thanh et al. [37]. In all the aforementioned contributions only static PDEs have
been considered. For time dependent problems, Raissi et al. [38] suggested to exploit the structure of Implicit
Runge–Kutta (IRK) time integrators. The IRK stages and the final solution are considered as output neurons of the
neural network.
Building on the aforementioned pioneering contributions, in this work the Neural Particle Method (NPM) is
presented as an innovative approach to solve the incompressible, inviscid Euler equations. In order to account for
large deformations and free surface flow, the Euler equations are formulated in an Updated Lagrangian frame.
Velocity and pressure are approximated with a neural network as a global ansatz function. Following Berg and
Nyström [35], the boundary conditions can be fulfilled exactly. The temporal discretization is realized with the high
order IRK integration scheme of Raissi et al. [38]. It will be demonstrated that the NPM fulfills the incompressibility
constraint without any stabilization even on arbitrarily distributed discretization points. This is a severe benefit of
the NPM compared to state of the art numerical methods.
The outline of this paper is as follows: First, the IRK method of Raissi et al. [38] is reviewed for second order
Ordinary Differential Equations (ODEs). The NPM is developed step by step in Section 3. In Section 4, the method
is validated and its performance illustrated by means of numerical examples of practical relevance, namely sloshing
in a container and the classical dam break test case. The paper concludes in Section 5.

2. A physics informed neural network for second order ODEs

Computational mechanics is about the differential formulation and solution of the equation of motion, which is
a second order PDE in time. In the presence of damping, the latter can be expressed in terms of the acceleration a
which is a function of time t, velocity v and position x:
a = a (v (x (t)) , x (t) , t) (1)
The idea of Runge–Kutta integration is to evaluate the acceleration a at a distinct number of stages s between
two time steps tn and tn+1 , i.e. at time instances tn + ci ∆t. The coefficients ci are given by the Butcher tableau
along with the coefficients a ji and b j appearing in (3) and (4). The indices i and j of the coefficient denote the
Runge–Kutta stage and range from one to the total number of stages s. Implicit Runge–Kutta (IRK) methods are
4 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

Fig. 1. Schematic of a one dimensional Physics Informed Neural Network adapted from [35]. The network inputs the known position xn
and outputs the current velocity vn+1 as well as the IRK velocity stages v j .

stable even when applied to stiff equations. Higher orders of accuracy can be achieved by increasing the number
of IRK stages s. Details on the theoretical background can be found in [39] or [40]. For the sake of brevity, the
following abbreviations for the position and velocity stages are introduced:

xi = x tn + ci ∆t , vi = v x tn + ci ∆t , tn + ci ∆t
( ) ( ( ) )
(2)

Additionally, the time dependency of a variable will be expressed by a subscript, e.g. vn = v (tn ). The velocity
stages and the current velocity then follow the update formula:
s
∑
v j = vn + ∆t a ji ai vi , xi , tn + ci ∆t
( )

i=1
s (3)
∑
b a v j , x j , tn + ci ∆t
j j
( )
vn+1 = vn + ∆t
j=1

The update of the position is directly computed from the velocity stages:
s
∑
x j = xn + ∆t a ji vi
i=1
s (4)
∑
j j
xn+1 = xn + ∆t b v
j=1

The update formulae (3) and (4) yield a fully coupled system of equations of the size of the number of IRK stages
s × s. It can be solved e.g. by means of the Newton–Raphson method. For standard mesh-based and mesh-free
methods this is usually too expensive. Hence, lower order time integrators such as the second order Newmark scheme
are preferred. When a neural network is used as global spatial ansatz, Raissi et al. [38] suggested to put a neural
network prior on the Runge–Kutta stages and the solution. This allows the use of very high order IRK methods.
For the IRK integration of the equation of motion (1), the neural network takes the position of a discretization
point xn as input and outputs its IRK velocity stages vi as well as its velocity vn+1 at the next time step. Note that
the position is solely determined by the velocity through (4) and does not require a separate network output. The
concept is briefly sketched in Fig. 1. The parameters of the neural network are its weights wljk and its biases blj . In
a Feed-Forward Neural Network (FFNN), information is passed in one direction from the input towards the output.
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 5

The feed-forward algorithm for computing the output vl for a given input xn is defined by:
vl = σ l z l
( )

z l = W l σ l−1 z l−1 + bl
( )

z l−1 = W l−1 σ l−2 z l−2 + bl

( )

.. (5)
.
z = W 2 σ 1 z 1 + b2
2
( )

z 1 = W 1 x n + b1
Here, σ l denotes the activation function which is the hyberbolic tangent throughout this work. The components of
the weight matrices W l are wljk . Further details can be found e.g. in [35]. After initialization, the network parameters,
i.e. the weights and the biases are optimized in a training loop to reduce a predefined objective loss function. In
the case of Physics Informed Neural Networks, the loss function is designed to incorporate the governing physics.
For this purpose, the velocity update (3) is rearranged such that it yields an expression for the known velocity at
the previous time step:
s
∑
vnj = v j − ∆t a ji ai vi , xi , tn + ci ∆t
( )

i=1
s (6)
∑
vs+1 b a v , x , tn + c ∆t
j j j j i
( )
n = vn+1 − ∆t
j=1

j
The velocity estimates vn and vs+1
n are computed from the velocity IRK stages and the final solution, respectively.
Therefore, the estimates are dependent on the network output. To optimize the latter, a loss function SS E v is
introduced as the sum of squared errors of the velocity estimates:
s
⏐v − vn ⏐2 + ⏐vs+1 − vn ⏐2
∑ ⏐ j ⏐ ⏐ ⏐
SS E v = n n (7)
j=1

The equation of motion is then solved by training the neural network, i.e. by adjusting its weights and its biases
such that the loss (7) reaches a minimum. Throughout this work, a combination of Adam Optimization [41] and
L-BFGS-B gradient descent [42] is employed for training.
To illustrate the potential and the accuracy of the presented IRK integration, the initial value problem of a one
dimensional mass–spring–damper system with mass m, stiffness k and damping coefficient d is considered. Written
in generalized coordinates q, the equation of motion and its analytical solution q (t) for the case of under-critical
damping are given by
( √ )
q̈ + 2Dω0 q̇ + ω02 q = 0, q (t) = q̂ e−Dω0 t cos ω0 1 − D 2 t (8)

with q̂ the initial amplitude at time t = 0. The eigen frequency ω0 and the attenuation factor D are defined as:
k d
ω02 = , D= √ (9)
m 2 cm
For a detailed derivation, the reader is referred to standard text books, e.g. [43]. In Fig. 2, the IRK integrated
equation of motion (8) is plotted together with its analytical solution as a function of time. The system considered
has unit mass, unit stiffness and the damping coefficient is d = 0.1 N/s. A total number of s = 8 IRK stages
and two hidden layers with 20 neurons each have been used in all simulations. For training, 100 iterations with
an Adam Optimizer (learning rate 0.001) were followed by L-BFGS-B gradient descent until convergence. Fig. 2
neatly illustrates the potential of high order IRK integration for computational mechanics. Time steps can be chosen
so large that entire vibration cycles are spanned without losing accuracy. This is in sharp contrast to explicit time
integration schemes that are usually employed in meshfree particle methods.
6 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

Fig. 2. Comparison of Implicit Runge–Kutta integration and analytical solution of the 1D equation of motion (8) with m = 1 kg, k = 1 N/m
and d = 0.1 N/s. For all simulations, the network layout consists of one input neuron with 2 hidden layers of 20 neurons each and 9 output
neurons, i.e. s = 8 IRK stages. With this setup, large time steps that span entire oscillations cycles can be accurately realized.

3. Updated Lagrangian formulation

So far, only the solution of an ODE has been discussed. The mass–spring–damper system from the previous
section was considered as rigid. In CFD however, one is interested in the deformation of fluids which brings into
play spatial derivative operators. Inviscid incompressible fluids are described by the Euler equations. The latter
comprise the momentum equation and the balance of mass which requires the divergence of the velocity to vanish:
1
a=− grad p + b
ρ (10)
div v = 0
Here, ρ denotes the (constant) density, p the pressure and b the gravitational acceleration. The structure of the Euler
equations imposes difficulties for its numerical solution with traditional numerical methods. To avoid instabilities
arising from the incompressibility constraint, the original equations are often modified using stabilizing terms as
briefly discussed in Section 1. In contrast, it will be demonstrated that the neural network architecture presented in
this work is able to compute the pressure that fulfills the incompressibility constraint exactly without any additional
algorithmic treatment.
In order to evolve (10) in time, a neural network is constructed that inputs analogously to Section 2 the position
xn at time tn . For each position xn , the network is trained to predict the IRK velocity stages vi , the velocity of the
next time step vn+1 and the pressure stages pi . Due to this network architecture, spatial derivatives can only be
constructed with respect to the previous time step tn . Therefore, an Updated Lagrangian formulation is employed in
which the configuration at time tn is considered as the reference configuration. Spatial derivatives are computed with
respect to the reference configuration and are then transformed into the current one via a push-forward operation.
The mapping between both configurations is defined by the incremental deformation gradient ∆F:
∂• ∂• ∂xn ∂• ∂xn+1
≈ · = · (∆Fn+1 )−1 , ∆Fn+1 = (11)
∂xn+1 ∂xn ∂xn+1 ∂xn ∂xn
Note that this procedure corresponds to the chain rule of differentiation. In order to apply the IRK integration
introduced in the previous section, the Euler equations must be evaluated at each IRK stage. This requires the
computation of an incremental deformation gradient from each position stage xi . Since the neural network outputs
velocity stages vi , first the rate of the incremental deformation gradient ∆Ḟ is computed. Making use of (4), the
incremental deformation gradient itself can then be obtained from its rate through IRK integration:
s s
∂vi ∂xi ∑ ∂vi ∑
∆Ḟi = , ∆Fi = = 1 + ∆t a ji = 1 + ∆t a ji ∆Ḟi
∂xn ∂xn i=1
∂x n i=1
s s (12)
∂vn+1 ∂xn+1 ∑ ∂v j ∑
∆Ḟn+1 = , ∆Fn+1 = = 1 + ∆t bj = 1 + ∆t b j ∆Ḟ j
∂xn ∂xn j=1
∂xn j=1
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 7

Using the above defined strain measures, the velocity divergence at each IRK stage and at the next time step is
computed from:
∂vi [ )−1 ]
div vi = tr i = tr ∆Ḟi · ∆Fi
(
∂x (13)
∂vn+1
= tr ∆Ḟn+1 · (∆Fn+1 )−1
[ ]
div vn+1 = tr
∂xn+1
The gradient of the IRK pressure stages pi is computed from:
∂ pi ∂ pi
grad pi = = · (∆F)−1 (14)
∂xi ∂xn
An evolution equation for the pressure does not exist. While one could alternatively derive a pressure Poisson
equation from the divergence of the momentum equation [44], this was not found to be necessary within the NPM.
The pressure is rather computed for each IRK stage such that the incompressibility constraint is fulfilled. Therefore,
only the pressure stages are considered as output neurons while the current pressure is defined as the weighted
average of pressure stages p j :
s
∑ s
∑
pn+1 = b p ,
j j
with bj = 1 (15)
j=1 j=1

The current velocity vn+1 is obtained directly from the output of the neural network while the update of spatial
coordinates follows (4). In order to account for the mass equation, the velocity divergence is added to the loss
function (7) as the sum of squared errors:
s
⏐div vi ⏐2 + |div vn+1 |2
∑ ⏐ ⏐
SS E div v = (16)
i=1
Following the
( original
) paper of Raissi et al. [38], Dirichlet type boundary conditions for the velocity v (x̄v ) = v̄ and
pressure p x̄ p = p̄ must also be accounted for in the loss function. Their contributions from the sum of squared
errors are:
s s
⏐2
⏐ p x̄ p − p̄ ⏐2
∑ ∑
v (x̄v ) − v̄ + |vn+1 (x̄v ) − v̄| ,
⏐ i 2
⏐ i( ) ⏐
SS E v̄ = ⏐ ⏐ SS E p̄ = (17)
i=1 i=1
Note that the contributions SS E v (7) and SS E div v (16) involve a summation over all points and the boundary
contributions (17) a summation over all boundary points. These have been omitted for brevity. The original treatment
of boundary conditions suggested by Raissi et al. [38] according to (17) is critically discussed in the remainder of
this section. In addition, an approach to account for contact with rigid walls is presented.

3.1. Imposition of Dirichlet boundaries

Raissi et al. [38] suggest the simultaneous training of boundary data and PDE. However, when simulating multiple
time steps, the accuracy of this approach is not sufficient. This can be illustrated by a simple example. A container of
unit width and height filled with an incompressible fluid of unit density under constant gravity acceleration g = 10
m/s2 is loaded only by its own weight. As Dirichlet conditions, the velocity normal to the side and bottom walls
must be zero as well as the pressure on the top of the domain. An accurate dynamic simulation must maintain both a
linear static pressure field and the boundary conditions over time. As shown in Fig. 3, the original method of Raissi
et al. [38] fails to meet this requirement. The fluid leaks at the boundaries. In [35], an alternative formulation for
imposing boundary conditions is suggested. The ansatz for a scalar valued primary variable u (x) can be written as
the sum of a smooth extension of the boundary data G (x) and a smooth distance function D (x) multiplied with
the output of a neural network û (x):
u (x) = G (x) + D (x) û (x) (18)
Any loss function is then formulated in terms of the ansatz u (x) that fulfills the boundary conditions by definition.
Training of û (x) is then only necessary to compute the PDE inside the domain. If for complex geometries D (x) and
8 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

Fig. 3. Static pressure of an incompressible fluid in a container after 50 time steps (T = 50 s, ∆t = 1 s) discretized by 900 fluid particles.
Left: The Dirichlet boundary conditions are not fulfilled when simultaneously trained with the incompressible Euler equations. Right: Using
a velocity projection introduced by Berg and Nyström [35], the problem is solved exactly even on arbitrarily distributed discretization points
(empty circles). The pressure is projected onto a background mesh using the linear interpolation method scipy.interpolate.griddata()
in python 3.

G (x) are difficult to define analytically, both functions can be computed using low-capacity ANNs, which ensures
smoothness and differentiability. For simple rectangular geometries an analytical expression is preferred in order to
save computational resources. In the static pressure test case, the normal velocity at the walls is zero and therefore
the boundary extension G (x) vanishes. The velocity projection simplifies to
vx (x ) vx (xn+1 ) Dvx (xn ) Dvx (xn ) v̂x (x ) v̂x (xn+1 )
( i ( i) ) ( ) ( i ( i) )
= ◦ i i (19)
v iy xi v y (xn+1 ) Dv y (yn ) Dv y (yn ) v̂ y x v̂ y (xn+1 )
where ◦ denotes the Hadamard product, i.e. element wise multiplication. Since only first order spatial derivatives
appear in the Euler equations (10), it is sufficient that the distance function is C1 continuous. With w the width
and h the height of the container, the analytical distance functions Dvx (xn ) and Dv y (yn ) introduced above for the
static pressure test case can be defined as:
4 4 1
Dvx (xn ) = − 2 xn2 + xn , Dv y (yn ) = yn (20)
w w h
Note that herein the distance functions Dvx and Dv y are computed in terms of the position xn at time tn , since the a
priori computation of distance functions depending on the updated position is not possible. It will be demonstrated
that this approach does not alter the solution at tn+1 . As a result of the projection (19), the rate of the incremental
deformation gradient ∆Ḟ is now subject to the product rule of differentiation:
∂vi ∂D vx vxi ∂ v̂i Dvx (xn )
( i ) ( ) ( )
Dv x Dv x
i
∆Ḟ = = ◦ i + ◦ , D= (21)
∂xn ∂xn v y v iy Dv y Dv y ∂xn Dv y (yn )
Based on the redefined rate of the incremental deformation gradient ∆Ḟi (21), the incremental deformation gradient
(12), the velocity divergence (13) and the gradient of the pressure (14) are computed. The loss function reduces to
the contributions from the IRK integration of the velocity (7), the incompressibility constraint (16) and the pressure
Dirichlet boundary conditions (17)2 :
L = SS E v + SS E div v + SS E p̄ (22)
The static pressure field obtained with the boundary projection is shown in Fig. 3. Using the developed approach,
the exact pressure is accurately computed even on a highly irregular spatial discretization. This is a significant
advantage over traditional mesh-free methods.

3.2. Contact with rigid walls

The velocity projection introduced in the preceding section ensures the fulfillment of Dirichlet boundary
conditions for particles that belong to the boundary. However, it does not prevent fluid particles from crossing
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 9

Fig. 4. Setup of the sloshing test case. In the present work, the geometry is defined by w = h = 1 m and a = 0.01 m. The density of the
liquid is ρ = 1 kg/m3 and the gravity acceleration g = 1 m/s2 .

the boundary in the case of large deformations as occurring in the dam break example of Section 4.2. An additional
algorithm to treat the contact of the fluid with a rigid wall is required. We choose the penalty approach which can
be regarded as an artificial volumetric spring force fic . The elongation of the spring is the signed gap g i between
the prescribed boundary position x̄ and the predicted current IRK stages of the position xi . It is penalized by the
parameter εc :
fic = εc g i a g i n, g i = xi − x̄ · n
( ) ( )
(23)
The vector n is the outer surface normal.
( ) The spring is only active when a particle has crossed a boundary. This is
modeled by an activation function a g i which takes values in the range from 0 (no contact) to 1 (contact):
⎧
⎪0 if g i < 0
( i) ( ( i )) ⎨
a g = 0.5 1 + sign g = 0.5 if g i = 0 (24)
if g > 0
⎪ i
1
⎩

For further information on the penalty method and contact mechanics, the reader is referred to standard text books,
e.g. [45].

4. Numerical results
The performance of the Neural Particle Method is illustrated by means of numerical examples. To demonstrate
its excellent conservation properties, free sloshing oscillations in a container are simulated for the inviscid case.
Both small and large amplitude sloshing can be displayed. The classical dam break test case is simulated until the
fluid hits the opposite wall. The simulations are in excellent agreement with experimental results reported in the
literature.

4.1. Inviscid free sloshing

Sloshing in an open container of unit width w and unit height h is considered. The fluid has unit density and
is subject to gravity of a unit magnitude. Initially, the surface elevation η of the fluid follows a sine profile with
amplitude a:
(π ( w ))
η (x) = h − a sin x− (25)
w 2
The normal velocity to the side and bottom walls is equal to zero as well as the pressure on the fluid free surface. The
geometry and the boundary conditions are graphically summarized in Fig. 4. This test case is especially interesting
because for small sloshing amplitudes an analytical solution exists. It is widely used as benchmark test case in
the literature, see e.g. [6,46,47] or [48]. In order to evaluate the conservation properties of the method, the total
10 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

Fig. 5. Constant network layout – variable time step: Amplitude–time diagram for the inviscid sloshing test case with different time steps.
The network layout used in all simulations consists of 2 input neurons, 2 hidden layers with 60 neurons each and 62 output neurons, i.e.
q = 20 IRK stages. This corresponds to the layout 1 from Table 1.

Fig. 6. Constant network layout – variable time step: Energy–time diagram (left) and kinetic energy–time diagram (right) for the inviscid
sloshing test case with different time steps. The network layout is the same as in Fig. 5.

energy of the system is observed. The specific total energy is comprised of a pressure, a kinetic and a potential
contribution:
1
E tot = p + ρ v2 + ρ g h (26)
2
In order to study the conservation properties of NPM, small amplitudes a = 0.01 m are considered first. Unless
otherwise stated, the spatial discretization consists of 900 nearly equi-spaced points where the height coordinate of
all inner and surface particles is shifted according to the surface elevation (25).

Constant network layout – variable time step

To investigate the convergence of the method with respect to the temporal discretization, the network layout is
kept constant. It consists of 2 input neurons, 2 hidden layers with 60 neurons each and 62 output neurons, i.e. s = 20
IRK stages. The first 14 sloshing amplitudes for the inviscid case are plotted in Fig. 5 for different time step sizes
∆t ∈ [0.1, 1] s. When looking at the energy–time diagram in Fig. 6, it is observed that the energy is only conserved
for the small time step ∆t = 0.1 s. For a large time step ∆t = 1 s, artificial damping becomes dominant and the
amplitude decays rapidly. Whilst the variation in pressure and potential energy is below 1% in all cases, the kinetic
energy is significantly decaying. Next, it is examined whether the restrictions on the time step size can be relaxed
if more hidden neurons are added to the network.
Constant time step – variable network layout
Different network architectures summarized in Table 1 have been examined in combination with a large time
step ∆t = 1 s. The computed amplitude as well as the kinetic energy of the system are plotted in Fig. 7. It is
found that increasing the number of IRK stages and the complexity of the neural network does not overcome the
restrictions on the time step size. This result is in sharp contrast to the findings of Raissi et al. [38], who presented
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 11

Fig. 7. Constant time step – variable network layout: Amplitude–time diagram (left) and kinetic energy–time diagram (right) for the inviscid
sloshing test case with different number of IRK stages for a large time step ∆t = 1 s. The labels correspond to different network layouts
summarized in Table 1.

Table 1
List of network layouts used for the results presented in Fig. 7.
Label Layout IRK stages
1 [2, 60, 60, 62] 20
2 [2, 100, 100, 100, 100, 100, 100, 152] 50
3 [2, 200, 200, 200, 152] 50
4 [2, 300, 300, 300, 300, 300, 300, 152] 50

the IRK integration method as a versatile tool enabling the use of arbitrarily large time steps. The limitation is
likely due to the underlying assumption of the Updated Lagrangian formulation on the spatial derivatives, namely
the push-forward operation using the incremental deformation gradient:
∂• ∂• ∂xn
≈ ∆F−1 , ∆F = (27)
∂xn+1 ∂xn ∂xn+1
Note that the deformation gradient is a linear operator. It can be derived from a Taylor series expansion of the
displacement, see e.g. [49]. The current position can be expressed in terms of the position at time tn and the
displacement increment between both configurations as ∆un+1 = un+1 − un . In differential form this yields:

dxn+1 = dxn + d∆un+1

∂∆u (xn ) ∂ 2 u (xn )
( )
= 1+ dxn + dxn ⊗ dxn + · · ·
∂xn ∂x2n (28)
∂ 2 ∆u (xn )
= ∆F dxn + dxn ⊗ dxn + · · ·
∂x2n
While generally higher order approximations are possible, these necessitate a meaningful definition of the line
increment dxn which is beyond the scope of the present work.

Sloshing at large amplitudes

The sloshing amplitude has been increased about a factor of twenty to a = 0.2 m. The spatial discretization
consists of 100 fluid particles at each fluid boundary and 700 interior particles. All points are randomly distributed.
The network layout 1 from Table 1 and a time step ∆t = 0.1 s are employed. For such a large amplitude, additional
waves are overlapping causing the wave to break. The numerical results are displayed in Fig. 8. They are in good
agreement with the results reported by Oñate et al. [48] obtained with the Particle Finite Element Method.
12 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

Fig. 8. Inviscid sloshing with high amplitude a = 0.2 m at T = 0 s, 1.9 s, 3.7 s, 5.7 s (from left to right).

Fig. 9. Dam break simulation setup.

Source: Adapted from [50].

Table 2
Parameter used in the dam break problem.
Parameter Symbol Value
Gravity acceleration g 9.8 m/s2
Length L 0.146 m
Penalty parameter εc 107

4.2. Dam break

The dam break example is a classical validation test case for Lagrangian fluid simulations, see e.g. [4,50] or [51].
Tabulated experimental data has been reported by Martin and Moyce [52]. The geometry is sketched in Fig. 9 and the
material data tabulated in Table 2. Note that in this paper only the outflow of the water column has been simulated.
In order to simulate the fluids interaction with an opposite wall, an additional surface identification algorithm would
be required to ensure a correct imposition of the zero pressure boundary condition. This may be realized for example
with the alpha-shape technique which is based upon a Delaunay triangulation of the domain, see [53]. However,
surface identification is beyond the scope of the present work and only the outflow of the fluid is considered.
Besides the pressure boundary, special attention must be paid at the velocity boundaries to ensure a proper
imposition of the slip condition. In SPH, dummy particles with zero velocity are used to enforce the boundary
constraints. With a neural network as global ansatz, this boundary treatment corresponds to a stick condition and is
not feasible. But also the velocity projection (20) would result in zero velocity at the container edges and result in
a stick condition. To allow the inviscid fluid to slip along the container walls, the velocity projection is relaxed. If
the distance of a fluid particle to an edge falls below a threshold, the particle is shifted along the wall in direction
of the fluid motion. It then becomes part of the boundary which is perpendicular to the original one. Note that a
linear velocity projection (20)2 is used in both spatial directions, where the width and the height of the domain are
computed prior to each time step.
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 13

Fig. 10. Comparison of measured and simulated evolution of the water front tip. The experimental data is taken from [52]. The training
set consists of 20 particles per unit L. No significant difference in results could be observed to the same simulations with 25 particles per
unit L.

Fig. 11. Simulation results at T = 0 s, 0.1 s, 0.2 s, 0.3 s (left to right). The results agree well with those reported by Koshizuka and Oka
[51]. The time step is ∆t = 0.01 and the discretization consists of 25 particles per unit L. Only a subset of the displayed points was used
for training.

Fig. 10 demonstrates an excellent agreement of the simulated fluid front tip evolution with the experimental
data reported by Martin and Moyce [52]. Network layout 1 from Table 1 has been employed in all simulations.
The results have been obtained on an equispaced discretization of 20 particles per unit L. A refinement to 25
particles per unit L did not affect the results. The overall shape of the collapsed water column in the time interval
T = [0, 0.3] s plotted in Fig. 11 also agrees well with the simulation results of Koshizuka and Oka [51]. Only
difference is the reduced wettability of the fluid near the vertical wall. This may be due to the imposition of the
zero pressure boundary condition on the free surface, which is not updated in the absence of a surface identification
algorithm as mentioned above.

5. Conclusion
The Neural Particle Method (NPM) has been proposed as a versatile simulation tool for incompressible fluid flow
involving free surfaces. A feed forward neural network is chosen as spatial ansatz function for the velocity and the
pressure [31]. Boundary conditions are exactly fulfilled due to the implementation of a boundary projection method
introduced by Berg and Nyström [35]. For the temporal integration, high order Implicit Runge–Kutta (IRK) methods
are applied [38]. The inviscid, incompressible Euler equations were formulated in an Updated Lagrangian manner.
NPM computes the pressure that accurately fulfills the incompressibility constraint while topological restrictions on
the discretization are not required. This superior behavior is in sharp contrast to state of the art numerical methods
which easily fail when it comes to disordered particle configurations. The conservation properties of the method
14 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127

were demonstrated in a sloshing test case. Additionally it was shown that the method performs well even for large
sloshing amplitudes.
Although high order IRK methods were used, the admissible time step size for the problems at hand is limited. In
the Updated Lagrangian framework, spatial derivatives are pushed into the current configuration with an incremental
deformation gradient. However, the latter is only a linear operator which restricts the time step size. In order to
exploit the full potential of the IRK integration, future work should focus on a relaxation of the time step constraint
originating from the linear deformation map.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could
have appeared to influence the work reported in this paper.

Data availability
The code for the numerical examples can be downloaded from https://gitlab.com/henningwessels/npm.

Acknowledgment
The first author wants to thank Jan Niklas Fuhg for the fruitful discussions.

References
[1] S.V. Patankar, Numerical heat transfer and fluid flow, in: Series in computational methods in mechanics and thermal sciences,
Hemisphere Publ. Co, New York, ISBN: 9780891165224, 1980.
[2] C.W. Hirt, B.D. Nichols, Volume of fluid (VOF) method for the dynamics of free boundaries, J. Comput. Phys. (ISSN: 00219991) 39
(1) (1981) 201–225, http://dx.doi.org/10.1016/0021-9991(81)90145-5.
[3] F. Brezzi, On the existence, uniqueness and approximation of saddle-point problems arising from Lagrangian multipliers, RAIRO Anal.
Numér. 8 (R2) (1974) 129–151, http://dx.doi.org/10.1051/m2an/197408R201291.
[4] C.W. Hirt, A.A. Amsden, J.L. Cook, An arbitrary Lagrangian-Eulerian computing method for all flow speeds, J. Comput. Phys. (ISSN:
00219991) 14 (3) (1974) 227–253, http://dx.doi.org/10.1016/0021-9991(74)90051-5.
[5] T.E. Tezduyar, S. Mittal, S.E. Ray, R. Shih, Incompressible flow computations with stabilized bilinear and linear equal-order-
interpolation velocity-pressure elements, Comput. Methods Appl. Mech. Engrg. (ISSN: 00457825) 95 (2) (1992) 221–242, http:
//dx.doi.org/10.1016/0045-7825(92)90141-6.
[6] H. Braess, P. Wriggers, Arbitrary Lagrangian Eulerian finite element analysis of free surface flow, Comput. Methods Appl. Mech.
Engrg. 190 (1–2) (2000) 95–109.
[7] L.B. Lucy, A numerical approach to the testing of the fission hypothesis, Astron. J. (ISSN: 00046256) 82 (1977) 1013, http:
//dx.doi.org/10.1086/112164.
[8] R.A. Gingold, J.J. Monaghan, Smoothed particle hydrodynamics: Theory and application to non-spherical stars, Mon. Not. R. Astron.
Soc. (ISSN: 0035-8711) 181 (3) (1977) 375–389, http://dx.doi.org/10.1093/mnras/181.3.375.
[9] W.K. Liu, S. Li, T. Belytschko, Moving least-square reproducing kernel methods (I) Methodology and convergence, Comput. Methods
Appl. Mech. Engrg. (ISSN: 00457825) 143 (1–2) (1997) 113–154, http://dx.doi.org/10.1016/S0045-7825(96)01132-2.
[10] S.J. Lind, R. Xu, P.K. Stansby, B.D. Rogers, Incompressible smoothed particle hydrodynamics for free-surface flows: a generalised
diffusion-based algorithm for stability and validations for impulsive flows and propagating waves, J. Comput. Phys. (ISSN: 00219991)
231 (4) (2012) 1499–1523, http://dx.doi.org/10.1016/j.jcp.2011.10.027.
[11] B. Li, F. Habbal, M. Ortiz, Optimal transportation meshfree approximation schemes for fluid and plastic flows, Internat. J. Numer.
Methods Engrg. (ISSN: 00295981) 83 (12) (2010) 1541–1579, http://dx.doi.org/10.1002/nme.2869.
[12] C. Weißenfels, P. Wriggers, Stabilization algorithm for the optimal transportation meshfree approximation scheme, Comput. Methods
Appl. Mech. Engrg. (ISSN: 00457825) 329 (2018) 421–443, http://dx.doi.org/10.1016/j.cma.2017.09.031.
[13] G.C. Ganzenmüller, S. Hiermaier, M. May, On the similarity of meshless discretizations of peridynamics and smooth-particle
hydrodynamics, Comput. Struct. (ISSN: 00457949) 150 (2015) 71–78, http://dx.doi.org/10.1016/j.compstruc.2014.12.011.
[14] S. Li, W.K. Liu, Meshfree Particle Methods, Corrected ed., Springer, Berlin, Heidelberg, New York, ISBN: 3540222561, 2007.
[15] L.P. Franca, T.J.R. Hughes, Two classes of mixed finite element methods, Comput. Methods Appl. Mech. Engrg. (ISSN: 00457825)
69 (1) (1988) 89–129, http://dx.doi.org/10.1016/0045-7825(88)90168-5.
[16] F. Brezzi, L.P. Franca, T.J.R. Hughes, A. Russo, Stabilization techniques and subgrid scales capturing, in: I.S. Duff (Ed.), The State
of the Art in Numerical Analysis, in: The Institute of Mathematics and its Applications Conference Series, Oxford University Press,
Oxford, ISBN: 0198500149, 1997, pp. 391–406.
[17] E. Oñate, J. García, A finite element method for fluid–structure interaction with surface waves using a finite calculus formulation,
Comput. Methods Appl. Mech. Engrg. (ISSN: 00457825) 191 (6–7) (2001) 635–660, http://dx.doi.org/10.1016/S0045-7825(01)00306-1.
[18] T.J.R. Hughes, G.R. Feijóo, L. Mazzei, J.-B. Quincy, The variational multiscale method—a paradigm for computational mechanics,
Comput. Methods Appl. Mech. Engrg. (ISSN: 00457825) 166 (1–2) (1998) 3–24, http://dx.doi.org/10.1016/S0045-7825(98)00079-6.
 H. Wessels, C. Weißenfels and P. Wriggers / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113127 15

[19] N.B. Erichson, L. Mathelin, Z. Yao, S.L. Brunton, M.W. Mahoney, J.N. Kutz, Shallow learning for fluid flow reconstruction with
limited sensors and limited data, 2019, arXiv preprint arXiv:1902.07358.
[20] M. Raissi, A. Yazdani, G.E. Karniadakis, Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations,
Science (ISSN: 0036-8075) 367 (6481) (2020) 1026–1030, http://dx.doi.org/10.1126/science.aaw4741, URL https://science.sciencemag.
org/content/367/6481/1026.
[21] J.N. Kutz, Deep learning in fluid dynamics, J. Fluid Mech. 814 (2017) 1–4.
[22] A. Chatterjee, An introduction to the proper orthogonal decomposition, Current Sci. (2000) 808–817.
[23] G.E. Hinton, R.R. Salakhutdinov, Reducing the dimensionality of data with neural networks, Science (ISSN: 0036-8075) 313 (5786)
(2006) 504–507, http://dx.doi.org/10.1126/science.1127647, URL https://science.sciencemag.org/content/313/5786/504.
[24] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8) (1997) 1735–1780.
[25] T.Q. Chen, Y. Rubanova, J. Bettencourt, D.K. Duvenaud, Neural ordinary differential equations, in: Advances in Neural Information
Processing Systems, 2018, pp. 6571–6583.
[26] R. Maulik, A. Mohan, B. Lusch, S. Madireddy, P. Balaprakash, D. Livescu, Time-series learning of latent-space dynamics for
reduced-order model closure, Physica D 405 (2020) 132368.
[27] S.L. Brunton, B.R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annu. Rev. Fluid Mech. 52 (2020) 477–508.
[28] A.T. Mohan, N. Lubbers, D. Livescu, M. Chertkov, Embedding hard physical constraints in neural network coarse-graining of 3d
turbulence, 2020, arXiv preprint arXiv:2002.00021.
[29] G. Cybenko, Approximation by superpositions of a sigmoidal function, Math. Control Signals Systems (ISSN: 0932-4194) 2 (4) (1989)
303–314, http://dx.doi.org/10.1007/BF02551274.
[30] M. Leshno, V.Y. Lin, A. Pinkus, S. Schocken, Multilayer feedforward networks with a nonpolynomial activation function can
approximate any function, Neural Netw. (ISSN: 08936080) 6 (6) (1993) 861–867, http://dx.doi.org/10.1016/S0893-6080(05)80131-5.
[31] I.E. Lagaris, A. Likas, D.I. Fotiadis, Artificial neural networks for solving ordinary and partial differential equations, IEEE Trans.
Neural Netw. (ISSN: 1045-9227) 9 (5) (1998) 987–1000, http://dx.doi.org/10.1109/72.712178.
[32] A. Griewank, Evaluating Derivatives: Principles and Techniques of Algorithmic Differentation, second ed., SIAM, Philadelphia,
Pennsylvania, ISBN: 9780898716597, 2008.
[33] J. Korelc, P. Wriggers, Automation of Finite Element Methods, Springer, Cham, Switzerland, ISBN: 978-3-319-39003-1, 2016.
[34] I.E. Lagaris, A.C. Likas, D.G. Papageorgiou, Neural-network methods for boundary value problems with irregular boundaries, IEEE
Trans. Neural Netw. (ISSN: 1045-9227) 11 (5) (2000) 1041–1049, http://dx.doi.org/10.1109/72.870037.
[35] J. Berg, K. Nyström, A unified deep artificial neural network approach to partial differential equations in complex geometries,
Neurocomputing (ISSN: 09252312) 317 (2018) 28–41, http://dx.doi.org/10.1016/j.neucom.2018.06.056.
[36] E. Samaniego, C. Anitescu, S. Goswami, V.M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang, T. Rabczuk, An energy approach to the
solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applications,
Comput. Methods Appl. Mech. Engrg. (ISSN: 00457825) 362 (2020) 112790, http://dx.doi.org/10.1016/j.cma.2019.112790.
[37] V.M. Nguyen-Thanh, X. Zhuang, T. Rabczuk, A deep energy method for finite deformation hyperelasticity, Eur. J. Mech. A Solids
(ISSN: 09977538) 80 (2020) 103874, http://dx.doi.org/10.1016/j.euromechsol.2019.103874.
[38] M. Raissi, P. Perdikaris, G.E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward
and inverse problems involving nonlinear partial differential equations, J. Comput. Phys. (ISSN: 00219991) 378 (2019) 686–707,
http://dx.doi.org/10.1016/j.jcp.2018.10.045.
[39] A. Iserles, A First Course in the Numerical Analysis of Differential Equations, second ed., in: Cambridge texts in applied mathematics,
vol. 44, Cambridge University Press, Cambridge, ISBN: 9780511995569, 2012.
[40] E. Hairer, G. Wanner, Stiff and differential-algebraic problems, second, revised ed., in: Springer series in computational mathematics,
vol. 14, Springer, Berlin, ISBN: 978-3-642-05220-0, 2010.
[41] D.P. Kingma, J. Ba, Adam: A method for stochastic optimization, 2014, arxiv:1412.6980v9.
[42] D.C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization, Math. Program. (ISSN: 0025-5610) 45 (1–3)
(1989) 503–528, http://dx.doi.org/10.1007/BF01589116.
[43] K. Magnus, K. Popp, W. Sextro, Schwingungen: Physikalische Grundlagen und mathematische Behandlung von Schwingungen; mit
68 Aufgaben mit Lösungen, ninth ed., in: Lehrbuch, Springer Vieweg, Wiesbaden, ISBN: 978-3-8348-2574-2, 2013.
[44] J.H. Ferziger, M. Perić, Computational Methods for Fluid Dynamics, third, revised ed., Springer, Berlin and New York, ISBN:
978-3-540-42074-3, 2002.
[45] P. Wriggers, Nonlinear Finite Element Methods, Springer, Berlin, Heidelberg, ISBN: 978-3-540-71000-4, 2008.
[46] B. Ramaswamy, Numerical simulation of unsteady viscous free surface flow, J. Comput. Phys. (ISSN: 00219991) 90 (2) (1990)
396–430, http://dx.doi.org/10.1016/0021-9991(90)90173-X.
[47] R. Radovitzky, M. Ortiz, Lagrangian finite element analysis of Newtonian fluid flows, Internat. J. Numer. Methods Engrg. (ISSN:
00295981) 43 (4) (1998) 607–619, http://dx.doi.org/10.1002/(SICI)1097-0207(19981030)43:4<607::AID-NME399>3.0.CO;2-N.
[48] E. Oñate, S.R. Idelsohn, F. Del Pin, R. Aubry, The particle finite element method: An overview, Int. J. Comput. Methods (ISSN:
0219-8762) 01 (02) (2004) 267–307, http://dx.doi.org/10.1142/S0219876204000204.
[49] A.J.M. Spencer, Continuum Mechanics, Dover, London, New York, Longman, 2004.
[50] B. Ramaswamy, M. Kawahara, Lagrangian finite element analysis applied to viscous free surface fluid flow, Internat. J. Numer. Methods
Fluids (ISSN: 0271-2091) 7 (9) (1987) 953–984, http://dx.doi.org/10.1002/fld.1650070906.
[51] S. Koshizuka, Y. Oka, Moving-particle semi-implicit method for fragmentation of incompressible fluid, Nucl. Sci. Eng. (ISSN:
0029-5639) 123 (3) (1996) 421–434, http://dx.doi.org/10.13182/NSE96-A24205.
[52] J.C. Martin, W.J. Moyce, Part IV. An experimental study of the collapse of liquid columns on a rigid horizontal plane, Phil. Trans.
R. Soc. A (ISSN: 0080-4614) 244 (882) (1952) 312–324, http://dx.doi.org/10.1098/rsta.1952.0006.
[53] H. Edelsbrunner, E.P. Mücke, Three-dimensional alpha shapes, ACM Trans. Graph. (ISSN: 0730-0301) 13 (1) (1994) 43–72,
http://dx.doi.org/10.1145/174462.156635.