AI-accelerated CFD simulation based on

OpenFOAM and CPU/GPU computing?

K. Rojek1[0000−0002−2635−7345] , R. Wyrzykowski2[0000−0003−1724−1786] , and

P. Gepner3[0000−0003−0004−1729]
1
Czestochowa University of Technology, Czestochowa, Poland
krojek@icis.pcz.pl
2
Czestochowa University of Technology, Czestochowa, Poland
roman@icis.pcz.pl
3
Warsaw University of Technology
pawel.gepner@gmail.com

Abstract. In this paper, we propose a method for accelerating CFD

(computational fluid dynamics) simulations by integrating a conventional
CFD solver with our AI module. The investigated phenomenon is respon-
sible for chemical mixing. The considered CFD simulations belong to a
group of steady-state simulations and utilize the MixIT tool, which is
based on the OpenFOAM toolbox. The proposed module is implemented
as a CNN (convolutional neural network) supervised learning algorithm.
Our method distributes the data by creating a separate AI sub-model for
each quantity of the simulated phenomenon. These sub-models can then
be pipelined during the inference stage to reduce the execution time or
called one-by-one to reduce memory requirements.
We examine the performance of the proposed method depending on the
usage of the CPU or GPU platforms. For test experiments with varying
quantities conditions, we achieve time-to-solution reductions around a
factor of 10. Comparing simulation results based on the histogram com-
parison method shows the average accuracy for all the quantities around
92%.

Keywords: AI acceleration for CFD · convolutional neural networks ·

chemical mixing · 3D grids· OpenFOAM · MixIT · CPU/GPU computing.

1 Introduction

Machine learning and artificial intelligence (AI) methods have become pervasive
in recent years due to numerous algorithmic advances and the accessibility of
computational power [1, 7]. In computational fluid dynamics (CFD), these meth-
ods have been used to replace, accelerate or enhance existing solvers [13]. In this
?
The authors are grateful to the byteLAKE company for their substantive support.
We also thank Valerio Rizzo and Robert Daigle from Lenovo Data Center and An-
drzej Jankowski from Intel for their support.

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
2 K. Rojek et al.

work, we focus on the AI-based acceleration of a CFD tool used for chemical
mixing simulations.
Chemical mixing is a critical process used in various industries, such as phar-
maceutical, cosmetic, food, mineral, and plastic ones. It can include dry blend-
ing, emulsification, particle size reduction, paste mixing, and homogenization to
achieve your desired custom blend [6].
We propose a collection of domain-specific AI models and a method of in-
tegrating them with the stirred tank mixing analysis tool called MixIT. MixIT
[11] provides deep insights solutions to solve scale-up and troubleshooting prob-
lems. The tool utilizes the OpenFOAM toolbox [14] for meshing, simulation, and
data generation. It allows users to design, simulate and visualize phenomena of
chemical mixing. More detailed, MixIT provides geometry creation, performs
3-dimensional (3D) CFD flow simulations for stirred reactors, including tracer
simulations and heat transfer analysis. Moreover, it allows you to get perfor-
mance parameters, such as mixing intensity, power per unit volume, blend time,
critical suspension speed, gas hold-up, and mass transfer coefficients.
Our goal is to provide an interaction between AI and CFD solvers for much
faster analysis and reduced cost of trial & error experiments. The scope of
our research includes steady-state simulations, which use an iterative scheme
to progress to convergence. Steady-state models perform a mass and energy
balance of a process in an equilibrium state, independent of time [2]. In other
words, we assume that a solver calculates a set of iterations to achieve the con-
vergence state of the simulated phenomenon. Whence, our method is responsible
for predicting the convergence state with the AI models based on a few initial
iterations generated by the CFD solver. In this way, we do not need to calculate
intermediate iterations to produce the final result, so the time-to-solution is sig-
nificantly reduced. The proposed AI models make it possible to run many more
experiments and better explore the design space before decisions are made.
The contributions of this work are as follows:
– AI-based method that is integrated with a CFD solver and significantly
reduces the simulation process by predicting the convergence state of simu-
lation based on initial iterations generated by the CFD solver;
– method of AI integration with the MixIT tool that supports complex simu-
lations with size of ≈ 1 million cells based on the OpenFOAM toolbox and
high performance computing with both CPUs and graphic processing units
(GPUs);
– performance and accuracy analysis of the AI-accelerated simulations.

2 Related Work
Acceleration of CFD simulations is a long-standing problem in many application
domains, from industrial applications to fluid effects for computer graphics and
animation.
Many papers are focused on the adaptation of CFD codes to hardware archi-
tectures exploring modern compute accelerators such as GPU [12, 15, 16], Intel

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 3

Xeon Phi [20] or field-programmable gate array (FPGA) [17]. Building a simula-
tor can entail years of engineering effort and often must trade-off generality for
accuracy in a narrow range of settings. Among the main disadvantages of such
approaches are requirements of in-depth knowledge about complex and extensive
CFD codes, expensive and long-term process of portability across new hardware
platforms, and, as a result, relatively low-performance improvement compared
with the original CFD solver. In many cases, only a small kernel of the solver is
optimized.
Recent works have addressed increasing computational performance of CFD
simulations by implementing generalized AI models able to simulate various use
cases and geometries of simulations [10, 13]. It gives the opportunity of achieving
lower cost of trial & error experiments, faster prototyping, and parametrization.
Current AI frameworks support multiple computing platforms that provide code
portability with minimum additional effort.
More recently - and most related to this work - some authors have regarded
the fluid simulation process as a supervised regression problem. In [8], the au-
thors present a novel generative model to synthesize fluid simulations from a set
of reduced parameters. A convolutional neural network (CNN) is trained on a
collection of discrete, parameterizable fluid simulation velocity fields.
In work [22], J. Thompson et al. propose a data-driven approach that lever-
ages the approximation of deep learning to obtain fast and highly realistic sim-
ulations. They use a CNN with a highly tailored architecture to solve the linear
system. The authors rephrase the learning task as an unsupervised learning
problem. The key contribution is to incorporate loss training information from
multiple time-steps and perform various forms of data-augmentation.
In paper [8], the authors show that linear functions are less efficient than their
non-linear counterparts. In this sense, deep generative models implemented by
CNNs show promise for representing data in reduced dimensions due to their
capability to tailor non-linear functions to input data.
Work [10] introduces a machine learning framework for the acceleration of
Reynolds-averaged Navier-Stokes to predict steady-state turbulent eddy viscosi-
ties, given the initial conditions. As a result, they proposed a framework that is
hybridized with machine learning.
In [18], the authors present a general framework for learning simulation
and give a single model implementation that yields state-of-the-art performance
across a variety of challenging physical domains, involving fluids, rigid solids,
and deformable materials interacting with one another.
Our method for AI-accelerated CFD simulations is based on utilizing a set
of sub-models that are separately trained for each simulated quantity. This ap-
proach allows to reduce the memory requirements and operate on large CFD
meshes. The proposed data-driven approach provides a low entry barrier for
future researchers since the method can be easily tuned when the CFD solver
evolves.

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
4 K. Rojek et al.

3 Simulation of Chemical Mixing with MixIT tool

3.1 MixIT: simulation tool based on OpenFOAM
MixIT [11] is the next generation collaborative mixing analysis and scale-up
tool designed to facilitate comprehensive stirred tank analysis using lab and
plant data, empirical correlations, and advanced 3D CFD models. It combines
knowledge management techniques and mixing engineering (science) in a unified
environment deployable enterprise-wide.
This tool allows users to solve Euler-Lagrange simulations [9] and momentum
transfer from the bubbles to the liquid. The liquid flow is described with the
incompressible Reynolds-averaged Navier-Stokes equations using the standard
k- model.
The generation of 3D grids is performed with the OpenFOAM meshing tool
snappyHexMesh [9]. For Euler-Lagrange simulations, a separate grid for each
working volume is created using the preconfigured settings of MixIT. A mesh
of the complete domain is generated, and the working volume is defined by the
initial condition of the gas volume fraction with the OpenFOAM toolbox.

3.2 Using MixIT tool for simulation of chemical mixing

The chemical mixing simulation is based on the standard k- model. The goal is
to compute the converged state of the liquid mixture in a tank equipped with a
single impeller and a set of baffles (Fig. 1). Based on different settings of the input
parameters, we simulate a set of quantities, including the velocity vector field
U , pressure scalar field p, turbulent kinetic energy k of the substance, turbulent
dynamic viscosity mut, and turbulent kinetic energy dissipation rate .

Fig. 1: Scheme of the simulated phenomenon

To simplify the simulation process, we have selected a subset of parameters

responsible for the simulation flow. The CFD solver supports many scenarios;
however, our research includes three basic case studies:

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 5

– assuming the different liquid level of a mixed substance,

– exploring the full range of rotations per minutes (rpm) of the impeller,
– considering different viscosities of the mixed substance.

4 AI-based acceleration
4.1 Introduction of AI into simulation workflow
Conventional modeling with OpenFOAM involves multiple steps (Fig. 2a). The
first step includes pre-processing, where you need to create the geometry and
meshing. This step is often carried out with other tools. The next step is the
simulation. It is the part that we mainly focus on in this paper by providing the
AI-based acceleration. The third step is post-processing (visualization, result
analysis).
Our goal is to create solver-specific AI method to ensure the high accuracy of
predictions. Our approach belongs to a group of data-driven methods, where we
use partial results returned by the CFD solver. The advantage of this method is
that it does not require to take into account a complex structure of the simula-
tion, but focus on the data. Such an approach lowers the entry barrier for new
CFD adopters compared with other methods, such as a learning-aware approach
[5], which is based on the mathematical analysis of solver equations.
Fig. 2b presents the general scheme of the AI-accelerated simulation versus
the conventional non-AI simulation. It includes (i) the initial results computed
by the CFD solver and (ii) the AI-accelerated part executed by the proposed
AI module. The CFD solver produces results sequentially iteration by iteration,
where each iteration produces intermediate results of the simulation. All inter-
mediate results wrap up into what is called the simulation results. The proposed
method takes a set of initial iterations as an input, sends them to our AI mod-
ule, and generates the final iteration of the simulation. The AI module consists
of three stages: (i) data formatting and normalization, (ii) prediction with AI
model (inference), and (iii) data export.
Data formatting and normalization translate the data from the OpenFOAM
ASCII format to the set of arrays, where each array stores a respective quantity of

Fig. 2: Scheme of AI-accelerated simulation (b) versus conventional non-AI ap-

proach (a)

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
6 K. Rojek et al.

the simulation (U , p, , mut, and k). These quantities are normalized depending
on a user configuration. The linear normalization is performed based on the
following equation:

yi = xi /max(|maxV |, |minV |) · R, (1)

where yi is the normalized value, xi is the input value, maxV , minV are the
maximum and minimum values from all the initial iterations of a given quantity,
R is a radius value (in our experiments R = 1). When a dataset has a relatively
small value of median compared to the maximum value of the set (median value
is about 1% of the maximum), then we use a cube normalization with yi∗ = yi3 .
The AI-accelerated simulation is based on supervised learning, where a set of
initial iterations is taken as an input and returns the last iteration. For simulating
the selected phenomenon with MixIT and conventional non-AI approach, it is
required to execute 5000 iterations. At the same time, only the first NI iterations
create the initial iterations that produce input data for the AI module. Moreover,
to reduce the memory requirements, the initial dataset is composed of simulation
results corresponding to every SI -th iteration from the NI initial iterations. The
determination of parameters NI and SI is the subject of our future work. At
this moment, we have empirically selected them by training the AI model with
different values of the parameters, up to achieving an acceptable accuracy. For
the analyzed phenomenon, we set NI = 480, and SI = 20. As a result, we take the
initial data set composed from iterations 20, 40, 60, ..., 480, so 480/5000 = 0.096
of the CFD simulation has to be executed to start the inference.
The data export stage includes denormalization of the data and converting
them to the OpenFOAM ASCII format. This stage and data formatting one
are executed on the CPU, but the prediction with AI model (inference) can be
executed on either the CPU or GPU, depending on user preferences.

4.2 Idea of using AI for accelerating simulation

Our neural network is based on the ResNet network [3] organized as residual
blocks. In a network with residual blocks, each layer feeds into the next layer
and directly into the layers about two hops away. To handle relatively large
meshes (about 1 million cells), we have to reduce the original ResNet network
to 6 CNN layers.
To train the network, we use 90% of the total data set, referred to as the
training data. The remaining 10% are kept as the validation data for model
selection, allowing detection of any potential overfitting of the model.
Our AI model is responsible for getting results from 24 simulation iterations
(from iterations 20, 40, 60, ..., 480) as the input, feed the network, and return
the final iteration. Each iteration has a constant geometry and processes the
3D mesh with one million cells in our scenario. Moreover, we have five quanti-
ties that are taken as the input and returned as the output of the simulation.
One of the main challenges here is to store all those data in the memory during
the learning. To reduce memory requirements, we create a set of sub-models

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 7

that independently work on a single quantity. Thanks to this approach, all the
sub-models are learned sequentially, which significantly reduces memory require-
ments. This effect is especially important when the learning process is performed
on the GPU.
The proposed strategy also impacts the prediction (inference) part. Since we
have a set of sub-models, we can predict the result by calling each sub-model
one-by-one to reduce the memory requirements or perform pipeline predictions
for each quantity and improve the performance. The created pipelines simulta-
neously call all the sub-models, where each quantity is predicted independently.
In this way, our method can be executed on the GPU platform (in one-by-one
mode), or the CPU platform with a large amount of RAM (in a pipelined mode).

5 Experimental evaluation

5.1 Hardware and software platform

All experiments are performed on the Lenovo platform equipped with two Intel
Xeon Gold 6148 CPUs clocked at 2.40GHz and two NVIDIA V100 GPUs with
16GB of the global HBM2 memory. The host memory is 400GB.
For training the models, the CUDA framework v.10.1 with cuDNN v.7.6.5
is used. As a machine learning environment, we utilize TensorFlow v.2.3, the
Keras API v.2.4, and python v.3.8. The operating system is Ubuntu 20.04 LTS.
For the compilation of codes responsible for data formatting and export from/to
OpenFOAM, we use GCC v.9.3. The CFD simulations are executed using the
OpenFOAM toolbox v.1906 and MixIT v.4.2.

5.2 Performance and accuracy analysis

The first part of the analysis is focused on the accuracy results, while the second
one investigates the performance aspects. Since the accuracy evaluation for the
regression-based estimation is not so evident as for the classification method, we
have selected a set of metrics to validate the achieved results. In the beginning, we
compare the contour plots of each quantity created across the XZ cutting plane
defined in the center point of the impeller. The results are shown in Figs.3-5,
where we can see the converged states computed by the conventional CFD solver
(left side) and AI-accelerated approach (right side).
The obtained results show high similarity, especially for the values from the
upper bound of the range. Some quantities, such as k, and  have a meager
median value (below 1% of the maximum value). As a result, a relatively small
error of the prediction impacts the sharp shape of contours for the near-zero
values. The higher absolute values, the more smooth shape of the contour plot
can be observed.
To estimate the accuracy, we also use a set of statistical metrics. The first two
ones are correlation coefficients that measure the extent to which two variables
tend to change together. These coefficients describe both the strength and the

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
8 K. Rojek et al.

(a) CFD (b) AI

Fig. 3: Contour plot of the velocity magnitude vector field (U ) using either the
conventional CFD solver (a) or AI-accelerated approach (b)

(a) CFD (b) AI

Fig. 4: Contour plot of the pressure scalar field (p)

(a) CFD (b) AI

Fig. 5: Contour plot of turbulent kinetic energy dissipation rate ()

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 9

Table 1: Accuracy results with statistical metrics

Quantity Pearson’s coeff. Spearman’s coeff. RMSE Histogram equaliz. [%]
U 0.99 0.935 0.016 89.1
p 0.993 0.929 0.004 90.1
 0.983 0.973 0.023 90.3
k 0.943 0.934 0.036 99.4
mut 0.937 0.919 0.147 93.5
Average 0.969 0.938 0.045 92.5

direction of the relationship. Here, we use two coefficients, including the Pear-
son correlation that estimates the linear relationship between two continuous
variables, as well as the Spearman correlation that assesses the monotonic rela-
tionship between two continuous or ordinal variables. The Pearson correlation
varies from 0.93 for the mut quantity to 0.99 for U and p. The average Pearson
correlation for all the quantities is 0.97. It shows a high degree correlation be-
tween the CFD (computed by solver) and AI (predicted) values. The Spearman
correlation varies from 0.92 to 0.97 with the average value equal to 0.94. It shows
a strong monotonic association between the CFD and AI results.
The next statistical metric is the Root Mean Square Error (RMSE). It is
the standard deviation of the residuals (differences between the predicted and
observed values). Residuals are a measure of how far from the regression line
data points are. The implemented data normalization methods ensure that the
maximum distance from the X-axis is 1. RMSE varies from 0.004 for the p quan-
tity to 0.15 for the mut quantity. The average RMSE for all the quantities is
0.05. Based on these results, we can conclude that the proposed AI models are
well fit.
The last method of accuracy assessment is histogram comparison. In this
method, we create histograms for the CFD solver and AI module results and
estimate a numerical parameter that expresses how well two histograms match
with each other. The histogram comparison is made with the coefficient of de-
termination, which is the percentage of the variance of the dependent variable
predicted from the independent variable. The results vary from 89.1% to 99.4%,
with an average accuracy of 92.5%. All metrics are included in Table 1.
We have also performed a collective comparison of the results, where we plot
a function y(x), where x represents the results obtained from the CFD solver,
while y is the prediction. The results are shown in Fig. 6. The black line shows
the perfect similarity, while the blue dots reveal the prediction uncertainty.
Now we focus on the performance analysis. We start with comparing the
execution time for the AI module executed on the CPU and GPU. In this ex-
periment, the mesh size is one million cells, and the CFD solver is run only
on the CPU. The AI-accelerated part is executed in three steps, including data
formatting and data export (implemented on CPU), as well as the AI predic-
tion (performed on the CPU or GPU platform). This part includes 90.4% of the

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
10 K. Rojek et al.

Table 2: Comparison of execution time for CPU and GPU

CPU GPU Ratio GPU/CPU
Data formatting [s] 65.1
Prediction (inference) [s] 41.6 290.6 7.0
Data export [s] 9.5
Entire AI module [s] 116.2 365.2 3.1
Full simulation [s] 1483.3 1732.3 1.2

simulation. For the AI-accelerated approach, the full simulation includes 9.6%
of all CFD iterations executed by the CFD solver and the AI-accelerated part.
Data formatting takes 65.1 s, while the data export takes 9.5 s. The AI
prediction time depends on the selected platform, taking 41.6 s on the CPU and
290.6 s on the GPU platform. So, somewhat surprisingly, the AI-accelerated part
(formatting + prediction + export) runs 3.1 times faster on the CPU than in
the case when the the GPU is engaged. Considering the CFD solver overhead
(9.6% of all iterations), we can observe that this is the most time-consuming
component of the entire AI-accelerated simulation. So the final speedup of the
CPU-based simulation is 1.2 against the case when the GPU is engaged. The
performance details are summarized in Table 2, where the execution time for
the full simulation (last row) includes executing both the entire AI module and
the first 9.6% of the simulation, which takes 1367.2 s.
The reason for the performance loss for the GPU prediction (inference) is a
high time of data allocation on the GPU compared with CPU and multiple data
transfers between the host and GPU global memory. These transfers are required

Fig. 6: Comparison of simulation results for the conventional CFD solver and
AI-accelerated approach

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 11

Table 3: Comparison of execution time for simulation of chemical mixing using

either the conventional CFD solver or AI-accelerated approach
9.6% of sim. 90.4% of sim. 100% of sim.
CFD solver [s] 1367.2 12874.1 14241.2
AI-accelerated (CPU) [s] 1367.2 116.2 1483.3
Speedup 1 110.8 9.6
AI-accelerated (engaging GPU) [s] 1367.2 365.2 1732.3
Speedup 1 35.3 8.2

because each quantity has its sub-model that needs to be sequentially loaded.
On the CPU platform, all the predictions of sub-models can be pipelined, and
the memory overhead becomes much lower.
The final performance comparison considers the execution time for the con-
ventional CFD solver and AI-accelerated approach. The first 9.6% of the simula-
tion takes 1367.2 s. The remaining part takes 12874.1 s for the conventional CFD
solver. Using the AI-accelerated module, this part is computed 110.8 times faster
when executed on CPU, and 35.3 times faster when GPU is involved. As result,
the entire simulation time is reduced from 14241.2 s to 1483.3 s (9.6x speedup)
for the CPU, and to 1732.3 s (8.2x speedup) when the GPU is engaged. These
results are summarized in Table 3.
Fig. 7 illustrates the performance advantages of the proposed AI-accelerated
solution against the conventional CFD solver. Here the blue bars show the execu-
tion time for the whole simulation, while the orange ones correspond to executing
90.4% of the simulation. The two bars on the left side correspond to using exclu-
sively the conventional CFD solver, while other bars demonstrate the advantages
of using the AI module to reduce the execution time of the simulation. Fig. 7
not only illustrates the speedup achieved in this way but also demonstrates that
this speedup is finally limited by the time required to perform the initial 9.4%
of the simulation using the OpenFOAM CFD solver.

6 Conclusions and Future Works

The proposed approach to accelerate the CFD simulation of chemical mixing
allows us to reduce the simulation time by almost ten times compared to using
the conventional OpenFOAM CFD solver exclusively. The proposed AI module
uses 9.6% of the initial iterations of the solver and predicts the converged state
with 92.5% accuracy. It is expected that this reduction in the execution time
will translate [21] into decreasing the energy consumption significantly, which
means reducing the environmental footprint, including the carbon footprint [19].
However, the reliable evaluation of this effect is the subject of our future work
since it requires considering the whole workflow, including both the inference
and training stages.
Our method is fully integrated with the MixIT tool and supports 3D meshes
with one million cells. Thanks to a data-driven approach, this method does not

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
12 K. Rojek et al.

Fig. 7: Advantages in execution time (in seconds) achieved for the AI-accelerated
approach over the conventional CFD solver

require a high knowledge of the CFD solvers to integrate it with the proposed
solution. Such an integration gives a promising perspective to apply the method
for CFD solvers that constantly evolve since it does not require going deep into
CFD models.
The AI module is portable across the CPU and GPU platforms, which al-
lows us to utilize the GPU power in the training stage and provides high per-
formance of prediction using the CPU. The proposed pipelined model can sep-
arately train each quantity that significantly reduces memory requirements and
supports larger meshes on a single node platform.
Our method is still under development. Particular attention will be paid to
support more parameters of the CFD simulation of chemical mixing, including
shape, position and number of impellers, the shape of the tube, number of baf-
fles, and mesh geometry. Another direction of our research is providing further
accuracy improvement and reduce the number of initial iterations required by
the AI module.

References

1. Archibald, R. et al.: Integrating Deep Learning in Domain Sciences at Exascale.

arXiv preprint arXiv:2011.11188v1 (2020)
2. Bhatt, D., Zhang, B. W., Zuckerman, D. M.: Steady-state simulations using
weighted ensemble path sampling. The Journal of Chemical Physics, 133(1) (2010)

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
 AI-accelerated CFD simulation 13

3. Chen D., Hu F., Nian G., Yang T.: Deep Residual Learning for Nonlinear Regres-
sion. Entropy 22(2):193, (2020)
4. Ferziger, J.H., Peric, M., Street, R.L.: Computational Methods for Fluid Dynamics.
4th edn. Springer Nature Switzerland (2020)
5. Jim, M., Ray, D., Hesthaven, J.S., Rohde, C.: Constraint-Aware Neural Networks
for Riemann Problems. arXiv preprint arXiv:3327.50678 (2019)
6. Handbook of Industrial Mixing: Science and Practice. Paul, E.L., Atiemo-Obeng,
V., Kresta, S.M. (eds.), John Wiley & Sons, Hoboken NJ (2004)
7. Jouppi, N.P., Young, C., Patil, N., Patterson, D.: A Domain-Specific Architecture
for Deep Neural Networks. Commm. ACM, 61(9), 50–59 (2018)
8. Kim, B., et al.: Deep Fluids: A Generative Network for Parameterized Fluid Sim-
ulations. arXiv preprint arXiv:1806.02071v2 (2019)
9. Kreitmayer, D. et al.: CFD-based Characterization of the Single-use Bioreac-
tor XcellerexTM XDR-10 for Cell Culture Process Optimization. arXiv preprint
arXiv:3461.77983 (2020)
10. Maulik, R., Sharma, H., Patel, S., Lusch, B., Jennings, E.: Accelerating RANS
turbulence modeling using potential flow and machine learning. arXiv preprint
arXiv:1910.10878 (2019)
11. MixIT: the enterprise mixing analysis tool, https://mixing-solution.com/. Last ac-
cessed 5 Feb 2021
12. Mostafazadeh, B., Marti, F., Pourghassemi, B., Liu, F., Chandramowlish-
waran, A.: Unsteady Navier-Stokes Computations on GPU Architec-
tures. In: 23rd AIAA Computational Fluid Dynamics Conference (2017)
https://doi.org/10.2514/6.2017-4508
13. Obiols-Sales, O., Vishnu, A., Malaya, N., Chandramowlishwaran, A.: CFDNet: a
deep learning-based accelerator for fluid simulations. In: Proc. 34th ACM Int. Conf.
on Supercomputing (ICS’20), pp. 1-12, ACM (2020)
14. OpenFOAM: the open source CFD toolbox, https://www.openfoam.com. Last ac-
cessed 5 Feb 2021
15. Rojek, K. et al.: Adaptation of fluid model EULAG to graphics processing unit
architecture. Concurrency and Computation: Practice and Experience 27(4), 937–
957 (2015)
16. Rojek, K., Wyrzykowski, R., Kuczynski, L.: Systematic adaptation of stencil-based
3D MPDATA to GPU architectures. Concurrency and Computation: Practice and
Experience 29(9), e3970 (2017)
17. Rojek, K. Halbiniak, K., Kuczynski, L.: CFD code adaptation to the FPGA archi-
tecture. The International Journal of High Performance Computing Applications
35(1), 33–46 (2021)
18. Sanchez-Gonzalez, A. et al.: Learning to Simulate Complex Physics with Graph
Networks. arXiv preprint arXiv:3394.45567 (2020)
19. Schwartz, R., Dodge, J., Smith, N.A., Etzioni, O.: Green AI. Commm. ACM,
63(12), 54–63 (2020)
20. Szustak, L., Rojek, K., Olas, T., Kuczynski, L., Halbiniak, K., Gepner, P.: Adapta-
tion of MPDATA heterogeneous stencil computation to Intel Xeon Phi coprocessor.
Scientific Programming, 14 pages (2015) https://doi.org/10.1155/2015/642705
21. Szustak, L., Wyrzykowski, R. Olas, T. Mele, V.: Correlation of performance op-
timizations and energy consumption for stencil-based application on Intel Xeon
scalable processors. IEEE Trans. Parallel and Distributed Systems, 31(11), 2582–
2593 (2020)

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29
14 K. Rojek et al.

22. Tompson, J., Schlachter, K., Sprechmann, P., Perlin, K.H.: Accelerating Eulerian
Fluid Simulation with Convolutional Networks. In: ICML’17: Proc. of the 34th
International Conference on Machine Learning, PLMR 70, pp. 3424–3433 (2017)

ICCS Camera Ready Version 2021

To cite this paper please use the final published version:
DOI: 10.1007/978-3-030-77964-1_29