JOM, Vol. 76, No.

2, 2024
https://doi.org/10.1007/s11837-023-06325-0
Ó 2023 The Author(s)

TECHNICAL ARTICLE

Adsorbed Gas Storage Digital Twin

1,2
GEORG KLEPP

1.—Institute for Energyresearch (IFE), Technical University of Applied Science Ostwestfalen

Lippe (TH OWL), 32657 Lemgo, Germany. 2.—e-mail: georg.klepp@th-owl.de

One possibility for energy storage are fuels. With gaseous fuels like hydrogen
or methane, significant efforts are necessary for a feasible storage in terms of
compression or liquefaction. This is of particular importance in the mobility
sector. An alternative to high-pressure or cryogenic gas storage is the storage
by adsorption in porous media using nano-carbons, metal–organic frame-
works, or metal hydrides as adsorbents. In order to assess the performance of
the charging and discharging of adsorption tanks, the mass and energy bal-
ance as well as the phase equilibrium (adsorption isotherm) and, if present,
the spatial distribution of properties has to be considered. In order to simplify
the analysis and prediction of these models, an attempt is made to develop
digital twins based on machine learning. Neural networks and Gaussian
process regression are applied to replace the system of coupled nonlinear and
differential equations. The data basis used is generated by simulations. Thus,
it is possible to easily predict the performance of a storage tank for different
gases or to determine an optimum storage device (material selection and tank
design).

gases have to be compressed to very high pressures

INTRODUCTION
(up to 100 bar for methane and up to 700 bar for
With regard to the transformation of the energy hydrogen) or cooled to very low temperatures
system towards renewable energy sources, there is a ( 160°C for methane and 250°C for hydrogen).
transition from fossil fuels to green fuels. Gaseous The demand of the energy needed for compression
green fuels are mostly based on methane or hydro- and cooling is relatively high with regard to the
gen resulting from many different production pro- stored energy in the gaseous fuels. In addition, the
cesses. Electrical energy from wind or solar power demands on the storage equipment for very high
can be transformed and stored in chemical form as pressures and very low temperatures are high,
hydrogen or hydrocarbons. The resulting hydrogen leading to expensive storage vessels (i.e., type IV
may be used to enrich methane. Biogas from tanks for hydrogen storage and double-walled cry-
biomass or organic waste is a mixture of several otanks). These factors prohibit the propagation of
gases, with the main components methane and gaseous fuels.
carbon dioxide. Hydrogen may also be used to enrich A less expensive alternative is adsorption storage,
methane. using the adhesion of the gas particles on the wall of
For a successful transition, the use of these gases a porous material like nano-carbons, metal–organic
and gas mixtures as fuels has been investigated frameworks, or metal hydrides.1–6 Thus, in a tank
with regard to combustion and operation of furnaces with fixed volume at a given pressure and temper-
and engines. In addition, questions referring to the ature level, more gas can be stored together with the
storage of these gases and gas mixtures have to be porous adsorbent material.
addressed. For a high volumetric storage capacity of Much research7–13 has been carried out into
green gaseous fuels like methane and hydrogen, the appropriate adsorption material: on the one hand,
it should have a high volumetric and gravimetric
storage capacity, ideally at ambient conditions,
(Received August 22, 2023; accepted November 30, 2023;
published online December 19, 2023)
while on the other hand, it should be produced
readily and conveniently. The range of possible

951
952 Klepp

materials extends from reprocessed waste to use data-driven models which are computationally
designed structures in the nano-range. In addition less expensive than the physical models based on
to physisorptive bonding, chemisorption may also be nonlinear and differential equations. Here, we will
used to enhance the bonding. Here, the focus is on develop a machine learning model for the adsorp-
the application and modeling of physical adsorption. tion equilibrium and the mass end energy balance.
Due to the heat of adsorption, the temperature It might be used for the application of tanks with
decreased during charging and increases during nearly constant temperature and concentration
discharging, thus affecting the adsorption equilib- distribution in the adsorbent. It can also model the
riums: with decreasing temperature, the adsorption adsorption heat and mass transfer in a control
capacity decreases, while, with increasing temper- volume implemented in a CFD simulation. Espe-
ature, the adsorption capacity increases. Thus, the cially for the latter case, the iterative solution and
ability of quickly charging or discharging the tank is time consuming coupling of additional differential
influenced by the ability to adjust the temperature equations is replaced by the solution of algebraic
in the adsorbent. An optimization of the charging equations.
and discharging process includes the sizing of the Here, the feasibility of replacing the physical
tanks, the shape of the adsorbent material, addi- model of the adsorption with a data-driven model
tional cooling/heating of the tanks, adaption of the and its application to the charging and discharging
pressure level, and mass flow rates for single tank of gaseous fuels in adsorption tanks is investigated.
and tank bundles. The charging of a methane hydrogen mixture in
Adsorption gas storage is of particular interest for activated carbon (NORIT RGM1) has been chosen
applications with high volumetric storage capacity, as an example to implement this approach.
i.e., as fuel tanks in vehicles (e.g., ANG tank, First, the governing equations and the output of
Ingevity, North Charleston, SC, USA). Thus, some the computations are presented. Then, the methods
experience is already available concerning the stor- used for machine learning are described. Finally,
age of methane (natural gas) and hydrogen (for fuel the results are shown and the feasibility of this
cells).4–16 There are also seasonal energy storage approach is discussed.
applications (hydrogen stored in metal hydrides) for
domestic use (e.g., LAVO Energy Storage, Sydney, PHYSICAL MODEL
Australia or Picea, Home Power Solutions, Berlin,
A lumped parameter model for a cylindrical
Germany).
adsorption tank with area-averaged pressure and
The charging and discharging has been modeled
temperature values has been derived based on the
based on the adsorption equilibrium and the mass
mass and energy balance equations and an adsorp-
end energy balance. With regard to the distribution
tion isotherm model. This model is described in
of the concentration and temperature, lumped
detail in Ref. 19.
parameter models and computational fluid dynam-
The change in the gas mass stored in the tank
ics (CFD) models14–16 have been used. For simple
results from the amount in the gas phase and the
geometries and cooling by natural convection only,
adsorbed phase, with the adsorbent total porosity
the temperature and pressure distribution is
known. The gas density has been computed with the
roughly uniform. For real-life applications, the
ideal gas equation of state.
geometries are more complex, and the distribution
The inflow and outflow (mass flow rate) are
is not uniform. In order to reduce the time for
prescribed. Charging has been implemented by a
charging and discharging, temperature control may
positive mass flux and an increase in pressure,
be applied. With a large number of sorption and
discharging by a negative mass flux and decreasing
desorption cycles, the storage capacity deteriorates.
pressure.
With regard to the storage of gas mixtures, there
The changes of the averaged temperature are
are fewer data available.
given by the energy balance resulting from the heat
Hydrogen storage in a cylindrical reaction cham-
of adsorption and the heat flux to the ambient in
ber is described in Ref. 17 with an empirical model
dependence of the effective heat capacity, the heat
and a machine learning model based on data from a
transfer coefficient to the ambient, the exterior
numerical simulation. In Ref. 18, an adsorption
surface, the packing density, the heat of adsorption,
model based on machine learning methods success-
and the amount of adsorbed gas.
fully describes the adsorption behavior of supercrit-
The gas phase and the adsorbed phase are
ical CO2 and CH4 on three types of coal, based on
assumed to be in equilibrium and are described by
measured data. Thus, in addition to physical models
the Dubinin–Astakhov equation for the adsorption
and empirical correlations, machine learning mod-
isotherm. The adsorbed gas density was computed
els are a viable alternative to describe the adsorp-
from the boiling temperature and liquid density at
tion process.
boiling point and the thermal expansion of the
Physical adsorption models are complex, leading
liquefied gases. The adsorption potential was deter-
to expensive computations. Therefore, a more effi-
mined by the saturated vapor pressure calculated
cient and effective approach for design and opti-
from the critical values. The total volume adsorbed
mization purposes is desirable. One possibility is to
Adsorbed Gas Storage Digital Twin 953

was computed from the pure vapor adsorption We will investigate the prediction of the charging
volumes, assuming volume proportionality. The process of hydrogen, methane, carbon dioxide, car-
reference amount for each component was calcu- bon monoxide, and krypton in activated carbons in a
lated at the total pressure. cylindrical tank. The temperature, pressure, and
The data for the adsorbent (NORIT RGM1), are the mass of the adsorbed species have been com-
taken from the literature, as well as the data for the puted as a function of time, mass flow rate, maxi-
adsorbed gases and the property values.9 Mean mum storage pressure, and cylinder height and
values of the properties for the pressure and diameter.
temperature range considered in the simulation The methods used for machine learning are
have been used. artificial neural networks (ANN) and Gauss process
The differential equations were discretized with a regression (GPR), and the models have been used as
Euler-forward scheme. The time steps were implemented in Matlab.20
adjusted to ensure a steady change in the area- The datasets used for training, validation, and
averaged pressure. The step size was chosen to testing were generated with the method described
ensure the independence of the numerical solution in Ref. 9, and the parameter ranges are shown in
from the step size (average deviation smaller 1%). Table I. The ambient conditions were 300 K and
The adsorption equilibrium was solved at every 100 kPa. The parameters of the Dubinin–Astakhov
time step by iteration. In order to facilitate the equation are taken from the literature.21,22 As the
convergence and the computational effort, constant data points are generated for each time step, there
property values have been used. are significantly more data points in the time range
Some of the results for the charging of methane than for the other variable ranges. The value range
and hydrogen are shown in Figs. 1, 2, and 3. for NORIT RGM1pressure and mass span two
The computational costs are relatively high: the orders of magnitude. The data ranges of the vari-
step size is restricted due to the explicit computation ables characterizing the species have only five
(of particular importance with high storage pressures supporting points for the chosen species.
as for hydrogen) and the adsorption equilibrium was Neural networks are suitable for large datasets
solved by iteration. Thus, an alternative formulation, with no prior input about the underlying features.
which is computationally significantly less expen- They do not scale well for very large datasets. GPR
sive, is desirable. One possibility is to use a data- is therefore used for modeling smaller datasets with
driven machine learning model. some knowledge about the dependencies, and they
do scale well for very large datasets.
MACHINE LEARNING MODEL The dataset available has been divided into a
training and a validation set and, if possible, also a
The final target is to develop a machine learning
test set. The amount of data has to be sufficient.
model which can act as a digital twin of the
With the training/validation dataset, a five-fold
adsorption gas storage system. The significant
cross validation was performed.
amount of varied data can be computed by physical
models validated by experimental results. As a first
step, the feasibility of this approach has been
investigated and appropriate methods identified.

Fig. 1. Charging of cylindrical 33-L tank filled with nano-carbon with gas. Pressure (a) and temperature (b) distribution for hydrogen, methane,
carbon monoxide, carbon dioxide, and krypton at mass flux 3.9 10 3 kg/s; data from Ref. 19.
954 Klepp

Fig. 2. Charging of cylindrical 33L tank filled with nano-carbon with hydrogen. Pressure (a) and temperature (b) distribution for low mass flux
3.93 g/s and high mass flux 7.87 g/s; data from Ref. 19.

Fig. 3. Charging of cylindrical 33-LKe tank filled with activated carbon with hydrogen. Pressure (a) and temperature (b) distribution for 0.482 g/s
up to 35 bar, 3.85 g/s up to 350 bar, and 7.7 g/s up to 700 bar.

is being trained, the difference between predicted

Table I. Parameter ranges of design variables for values and actual values will be propagated back-
machine learning ward so that the weights and biases of each node can
be adjusted according to the error of that node.
Variable Range The size (number of neurons) and number of
hidden layers are varied, as well as the transfer
Maximum storage pressure 3.6 105–700 105 Pa
function. A shallow network is sufficient for this
Hydrogen concentration 0–1 g/g
Tank diameter 0.11–0.46 m problem. The transfer functions considered are the
Tank height 0.6–2.4 m rectified linear unit, hyperbolic tangent, and sig-
Mass flux 0.16–0.62 g/s moid. The best results were obtained here using a
rectified linear unit activation function. The Leven-
berg–Marquardt back propagation algorithm was
NEURAL NETWORKS used as a training function for the weights and
For the machine learning task, ANN were used. An biases. The neural networks were used as imple-
ANN is comprised of a network of artificial neurons mented in Matlab.20 The number of data needed to
(nodes) which are connected to each other. The input train and validate the neural network was
nodes take in information in a form which can be investigated.
numerically expressed. This information is then
passed through the network, through hidden layers, GAUSS PROCESS REGRESSION (GPR)
until it reaches the output nodes, then from node to GPR models have been widely used in machine
node, based on the connection strengths (weights and learning applications because of their representa-
biases) and the transfer functions. When the network tive flexibility and inherent uncertainty measures
Adsorbed Gas Storage Digital Twin 955

over predictions. GPR is a non-parametric approach teaching and validation datasets. The values of R2
that can be used for regression problems and is were
P computed P by Matlab as implemented, R2 = 1–
particularly suited for small datasets. It is based on (yi yi, Model)2/ (yi yMean 2
P ) as well as 2the root mean
the assumption that the output variable is a real- square error (RMSE) = ( (yi yi, Model) /n )1/2.
ization of a Gaussian process, which is a collection of Some of the results of the machine learning and
random variables, any finite number of which have testing are shown in Tables II and III. There are
a joint Gaussian distribution. First, a prior distri- only small differences with regard to the GPR
bution is specified, then the probabilities are relo- method used (see Table II). The ANN with 2 layers
cated based on the evidence (i.e., the observed data). of 10 neurons performs best of all the ANN, but still
Unknown points are predicted assuming a Gaussian predicts worse than the GPR models (see Table II).
distribution. The prior distribution is described by a The best results are for the GPR method with an
mean function and a covariance function. The forms exponential covariance function, so this ML model
of the mean function and the covariance kernel was chosen for the analysis. The training and
function in the GPR are chosen and tuned during validation data as well as the test data for the
model selection. The mean function chosen is con- pressure and temperature values were predicted
stant (mean of the training dataset) and the covari- with very good accuracy. Nevertheless, the test data
ance kernel function chosen is varied using the for the adsorbed mass was predicted with very poor
function implemented in Matlab.20 Here, the best accuracy by the all-species model (see Table III).
results were for the exponential kernel and the The prediction of the test data with a one-species
squared exponential kernel. model was fairly accurate, with R2 equal to unity.
Here, the amount and variety of the input data with
RESULTS AND DISCUSSION regard to the constants of the adsorption isotherm
model used in this analysis to generate the all-
For the modeling of the dependence on the
species model was not sufficient.
operating conditions and on the species concentra-
Using the one-species model, the charging of
tion, respectively, a dataset of 1747 9 11 points for
methane in a tank filled with activated carbon was
teaching and validation and a dataset of 1168 9 11
computed and compared with measured data from
points for testing were computed. Two types of
the literature15 (see Fig. 4). The predicted values
models were generated: models for single species
are slightly lower than the measured values, with
(methane, hydrogen, carbon dioxide) and one model
an average deviation for the computed pressure of
for all the species (methane, hydrogen, carbon
20% and, for the computed temperature, of 17%.
dioxide, carbon monoxide, krypton).
The deviations are higher for higher pressure
As input variables, time, maximum storage pres-
values, as the modeling error accumulates during
sure, mass flow, and tank diameter and height were
the charging process. The deviations of measured
used. The different gas species were characterized by
and computed data are mostly due to the simplifi-
their molecular weight, critical temperature, and
cations with regard to the time- and space-averaged
critical pressure. The output variables were pressure,
property values of the physical model used as basis
temperature, and adsorbed species concentration in
for the machine learning model.
the tank. Fivefold cross-validation was used for the

Table II. Correlation results of machine learning pressure distribution

GPR GPR GPR ANN ANN

Method exp. square exp. Matern 5/2 10 10/10

Computation time 25 s 20 s 21 s 7.9 s 16 s

Validation R2 1 1 1 1 1
Validation RMSE 11,500 15,800 12,000 306,000 19,600

Table III. Correlation results of machine leaning for the all-species-model, and different output variables

Variable Pressure Temperature Mass adsorbed

Computation time 25 s 32 s 9s
Validation R2 1 1 1
Validation RMSE 11,500 0.271 0.0128
Test R2 1 1 0
Test RMSE 12,06 0.330 0.218
956 Klepp

Fig. 4. Charging of a cylindrical 1.82-L tank filled with activated carbon with 1 L/min, 10 L/min, and 30 L/min methane. Comparison of computed
and measured values (data from Ref. 15) for pressure (a) and temperature (b).

CONCLUSION cluded in the article’s Creative Commons licence

and your intended use is not permitted by statutory
Machine learning models are a viable approach to
regulation or exceeds the permitted use, you will
describe adsorption phenomena The generated
need to obtain permission directly from the copy-
machine learning models may be used to predict right holder. To view a copy of this licence, visit h
the charging process of gases like methane, hydro-
ttp://creativecommons.org/licenses/by/4.0/.
gen, carbon dioxide, carbon monoxide, or krypton in
cylindrical tanks filled with activated carbon. The REFERENCES
influence of tank geometry and the temperature and 1. R.W. Judd, D.T.M. Gladding, R.C. Hodrien, D.R. Bates, J.P.
pressure transients during charging have been Ingram, and M. Allen, ACS Divis. Fuel Chem. 43, 10244
reproduced. Taking into account the distribution of (1998).
the property values will improve the prediction of 2. H. Li, K. Wang, Y. Sun, C.T. Lollar, J. Li, and H.-C. Zhou,
Mater. Today 21(2), 108 (2018).
experimental values. Most challenging is the mod- 3. D. DeSantis, J.A. Mason, B.D. James, C. Houchins, J.R.
eling of the adsorption characteristics and species Long, and M. Veenstra, Energy Fuels 31, 2024 (2017).
properties. Here, the datasets used for teaching and 4. S. Du, Y. Qu, H. Li, and X. Yu, Energies 15, 4261 https://doi.
validation should appropriately represent the value org/10.3390/en15124261 (2022).
range, without regard to actual real materials and 5. M. Nasser, T. Megahed, S. Ookaware, and H. Hassen, J.
Energy Syst. 6, 560 (2022).
species. 6. S. Rostami, A.N. Pour, and M. Izadyar, Sci. Progr. 101, 171
https://doi.org/10.3184/003685018X15173975498956 (2018).
FUNDING 7. A. Ertas, C.T.R. Boyce, and U. Gulbulak, Energies 13, 682 h
Open Access funding enabled and organized by ttps://doi.org/10.3390/en13030682 (2020).
8. C. Santos, F. Marcondes, and J.M. Gurgel, Appl. Therm.
Projekt DEAL. Eng. 29, 2365 (2009).
9. P. Pfeifer, R. Little, T. Rash, J. Romanos, and B. Maland,
Advanced Natural Gas Fuel Tank Project (California Energy
CONFLICT OF INTEREST Commission, Sacramento, 2017).
10. M. Prosniewski, T. Rash, J. Romanos, A. Gillespie, D. Stalla,
On behalf of all authors, the corresponding author E. Knight, A. Smith, and P. Pfeifer, Fuel 244, 447 https://doi.
states that there is no conflict of interest. org/10.1016/j.fuel.2019.02.022.7 (2019).
11. D. Nguyen, J. Kim, T. Nguyen, N. Kim, and H. Ahn, Appl.
Energy 310, 118552 (2022).
OPEN ACCESS 12. Y. Zhuo, S. Jung, and Y. Shen, Energy Fuels 35, 10908
(2021).
This article is licensed under a Creative Com- 13. J. Tan, M. Chai, K. He, and Y. Chen, Energies 15, 2673 h
mons Attribution 4.0 International License, which ttps://doi.org/10.3390/en15072673 (2022).
permits use, sharing, adaptation, distribution and 14. S. Sahoo and M. Ramgopal, Int. J. Petrochem. Sci. Eng. 2,
reproduction in any medium or format, as long as 00059 (2017).
15. P.K. Sahoo, M. John, B.L. Newalkar, N.V. Choudhary, and
you give appropriate credit to the original author(s) K.G. Ayappa, Ind. Eng. Chem. Res. 50(23), 1300 https://doi.
and the source, provide a link to the Creative org/10.1021/ie200241x (2011).
Commons licence, and indicate if changes were 16. P.K. Sahoo, B.P. Prajwal, S.K. Dasetty, M. John, B.L.
made. The images or other third party material in Newalkar, N.V. Choudary, and K.G. Ayappa, Appl. Energy
119, 190 (2014).
this article are included in the article’s Creative 17. C.S. Wang and J. Brinkerhoff, Int. J. Hydrogen Energy 46,
Commons licence, unless indicated otherwise in a 24256 https://doi.org/10.1016/j.ijhydene.2021.05.007 (2021).
credit line to the material. If material is not in-
Adsorbed Gas Storage Digital Twin 957

18. M. Meng, Z. Qiu, R. Zhong, Z. Liu, Y. Liu, and P. Chen, 21. H. Esfandian, N. Esfandian, and M. Rezazadeh, Int. J. Eng.
Chem. Eng. J. 368, 847 https://doi.org/10.1016/j.cej.2019.03. Trans. B Appl. 33(5), 712–719 https://doi.org/10.5829/ije.20
008 (2019). 20.33.05b.01 (2020).
19. G. Klepp, In 14th International Renewable Energy Storage 22. K. Kawazoe, T. Kawai, Y. Eguchi, and K. Itoga, J. Chem.
Conference 2020 (IRES 2020), 225 (2020) https://doi.org/10. Eng. Jpn. 7(3), 153 (1971).
2991/ahe.k.210202.033.
20. The Math Works, Inc. MATLAB. Version 2021b, Mas- Publisher’s Note Springer Nature remains neutral with re-
sachusetts.
gard to jurisdictional claims in published maps and institutional
affiliations.