Journal of The

Electrochemical Society

OPEN ACCESS You may also like

- Insight into LIB Diffusion Phenomena
Analysis of Li-Ion Battery Electrochemical Using Analytical Impedance Models
E. Woillez and M. Chandesris
Impedance Spectroscopy Data: An Easy-to- - Real-time Nonlinear Model Predictive
Control (NMPC) Strategies using Physics-
Implement Approach for Physics-Based Parameter Based Models for Advanced Lithium-ion
Battery Management System (BMS)
Estimation Using an Open-Source Tool Suryanarayana Kolluri, Sai Varun Aduru,
Manan Pathak et al.

- Hexic Polynomial Approximation for

To cite this article: Matthew D. Murbach and Daniel T. Schwartz 2018 J. Electrochem. Soc. 165 A297 Spherical Diffusion to Accelerate Accurate
Lithium-Ion Battery Simulation
Astitva Mishra, Venkat R. Subramanian
and Bharatkumar Suthar

View the article online for updates and enhancements.

This content was downloaded from IP address 42.60.248.215 on 10/11/2024 at 07:17

 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018) A297

Analysis of Li-Ion Battery Electrochemical Impedance

Spectroscopy Data: An Easy-to-Implement Approach for
Physics-Based Parameter Estimation Using an Open-Source Tool
∗
Matthew D. Murbach and Daniel T. Schwartz∗∗,z
Department of Chemical Engineering and Clean Energy Institute, University of Washington, Seattle,
Washington 98195-1750, USA

The quantitative analysis of electrochemical impedance spectroscopy (EIS) data is important for both characterization and prognostic
applications in many electrochemical systems. Here we describe an open-source platform, the ImpedanceAnalyzer, for easy-to-use
physics-based analysis of experimental EIS spectra. To demonstrate the use of the platform, we explore the basic capabilities of the
pseudo two-dimensional (P2D) battery model to predict publicly available experimental EIS data from a 1500 mAh commercial
lithium-ion (LiCoO2 /graphite) cell. An a priori computed dataset of 38,800 P2D-based impedance spectra simulations, covering a
wide range of frequencies (1 mHz to 100 kHz) and model parameters, enables a straightforward least squares matching approach
for analyzing experimental spectra. We find an average error of 1.73% between the best-matching computed spectrum from the
38,800 member library and the experimental spectrum being analyzed. Our analysis shows there is significant opportunity to improve
the fit between experimental data and physics-based impedance simulations by a combination of a larger computed dataset, local
optimization, and further additions to the model physics. The approach and open source tools developed here can be easily extended
to other electrochemical systems.
© The Author(s) 2018. Published by ECS. This is an open access article distributed under the terms of the Creative Commons
Attribution 4.0 License (CC BY, http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse of the work in any
medium, provided the original work is properly cited. [DOI: 10.1149/2.1021802jes]

Manuscript submitted November 9, 2017; revised manuscript received December 25, 2017. Published January 25, 2018.

Electrochemical impedance spectroscopy (EIS) is a powerful tool independent measurements to drastically narrow the number of fitted
for investigating a wide variety of electrochemical systems.1–3 EIS parameters.30 In short, today there is gap between the desire to use
spectra separate individual electrochemical processes by their charac- physics-based models to estimate parameters in EIS, and the actual
teristic timescales, enabling both qualitative and quantitative analysis (routine) use of physics-based models for parameter estimation from
of electron transport,4,5 reaction rates and mechanisms,6,7 intercala- data.
tion processes,8 mass transport,9,10 and electrode structure.11,12 The Here we demonstrate an easily implemented and extendable ap-
noninvasive nature of EIS also makes impedance measurements use- proach for leveraging sophisticated physics-based models of EIS spec-
ful in prognostic applications such as fuel cell health estimations13,14 tra to estimate parameters from experimental data. The parameter
or prediction of remaining useful lifetime in batteries.15,16 estimation approach used in this work relies on error minimization
Qualitative analysis of EIS spectra generally involves assessing the between experimental data and a large library of a priori simulated
shape of Nyquist plot features to determine the relative importance impedance spectra. One benefit of a dataset-based approach is that
of different physicochemical processes.17,18 In contrast, quantitative it always converges and the resulting parameter estimates are guar-
analysis relies on fitting a model to the data in order extract values for anteed to be reasonable if the original dataset was constructed from
specific thermodynamic, transport, and/or kinetic parameters. Most physically reasonable parameters. Another benefit of a dataset-based
experimental datasets are analyzed quantitatively using an equivalent approach is that global sensitivity analysis can be used to understand
circuit analog. Fitting an equivalent circuit to EIS data is straightfor- the contribution of different parameters to the model variance.31 On
ward using standard least squares regression techniques.19,20 A good the other hand, a dataset-based approach will generally not provide
fit can often be found with a relatively simple equivalent circuit, partic- the parameters with the lowest possible error between model and
ularly if non-ideal elements like the constant phase element are used. experiments.
Moreover, many simple equivalent circuits, like the Randles’ circuit,21 The work and software tool presented here is extendable to a
have physically interpretable parameters based on linearized electro- wide range of electrochemical systems, though our first implemen-
chemical processes. However, as more complex equivalent circuits are tation is using the Doyle-Fuller-Newman pseudo two-dimensional
derived and utilized, the lumped parameters can lose their direct phys- (P2D) lithium-ion battery (LIB) model as the basis for analyzing EIS
ical interpretability and the structure of the equivalent circuit analogs experiments.32,33 The earliest uses of LIB physics-based impedance
themselves can be degenerate.22 models have been to inform the analysis of experimental EIS data.
An alternative to equivalent circuits for quantitative analysis of EIS For example, Doyle et al.28 used physics-based EIS simulations to
data is to directly fit the data with a physics-based mathematical model show that the low frequency portion of a LIB impedance spectrum
of the electrochemical system. Many years of electrochemical model- appears (qualitatively) to be interpretable as a Warburg impedance,
ing research have laid the groundwork for the physics-based analysis but using a Warburg plot can produce erroneous (quantitative) dif-
of impedance in a wide variety of fields including corrosion,6 hy- fusivity estimates. Subsequently, additional physics has been added
drodynamic systems,23,24 fuel cells,25,26 and lithium-ion batteries.27–29 to the original P2D model to aid in the interpretation of EIS data.
Parameter estimation by fitting a physics-based model to experimen- For example, a surface oxide model was added to the positive elec-
tal EIS data is complicated by the fact that electrochemical models trode particles by Dees et al.30 to understand the increase in inter-
often contain a combination of coupled differential equations, alge- facial impedance with aging of LiNi0.8 Co0.15 Al0.05 O2 -based (NCA)
braic equations, and dozens of unknown parameters. For complex positive electrodes. Abraham et al.34 extended the model further to in-
physics-based models, convergence to a global best fit is rarely as- terpret the changing impedance response at different voltages in NCA
sured, even with an excellent initial guess for the unknown parameters. electrodes.
As a result, parameter estimation methods often need to rely on many Despite the widespread use of the P2D model for simulating LIBs
(2 of the top 4 most cited papers in the history of the Journal of the
∗ Electrochemical Society Student Member. Electrochemical Society (as of November 1st , 2017),35 the model has
∗∗ Electrochemical Society Fellow. seen limited use for quantitative analysis of experimental impedance
z
E-mail: [email protected] data. In this work, we present the ImpedanceAnalyzer, an open-source,
A298 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018)

web-based analysis platform aimed at making physics-based models diversity of lithium-ion battery chemistries, a dataset of different sim-
as easy to use as equivalent circuits for quantitative analysis of EIS ulations was synthesized using a wide, but physically meaningful,
experimental data. For the first implementation, we explore the basic range of parameters. The parameters and their ranges are shown in
capabilities of the original P2D impedance model to predict pub- Table AI in the Appendix. Parameters whose ranges spanned more
licly available experimental EIS data from a commercial cell. We than two orders of magnitude were sampled logarithmically. Sam-
discuss the implications of this work in terms of the next steps for pling the parameter space of the several dozen inputs to the P2D
expanding the model dataset, extending the models physics (surface model necessitates a large number of simulations. The Sobol’ sam-
oxide layers, surface-electrolyte interphase (SEI) layers, etc.), and pling sequence36 was used to efficiently explore the high dimensional
adding gradient-based parameter estimation to the basic Sobol’ pa- space and the SALib38 and savvy39 python packages were used to
rameter sampling method36 that forms the initial backbone of the generate the sampling sequences.
ImpedanceAnalyzer. As an open source tool, any of these modifi- An important, but subtle, issue in establishing model parameters
cations can be added by other researchers to improve performance. is that impedance measurements represent a “local” probe of the bat-
Overall, we believe the open source tools and approach presented here tery at the state-of-charge being assessed. The most relevant case
have the capability to combine knowledge generated from decades of for understanding the meaning of “local” is the treatment of electrode
physics-based impedance modeling research with the deep materials thermodynamics. The governing equations for the linear electrochem-
and chemistry insight of experimentalists to accelerate progress in the ical impedance (or the weakly nonlinear electrochemical impedance,
field. for that matter37 ) do not rely on knowledge of “global” relationship
between open circuit voltage and intercalated lithium concentration.
Instead, what matters is the local gradient. Thus, Table AI shows the
Methods
gradient of open circuit potential with intercalated concentration as a
The pseudo two-dimensional (P2D) lithium ion battery model.— parameter influencing the impedance. The local open circuit voltage
We have previously detailed the isothermal P2D model and efficient gradient values for each electrode are constrained by realistic bounds
frequency domain computational approach applied here to prepare found in the literature for multiple chemistries. As local variables, the
a library of simulated lithium-ion battery EIS spectra.37 Briefly, the thermodynamic gradients in each electrode are taken as independent
P2D model is a set of partial differential equations describing the parameters to be sampled like all others in Table AI.
one-dimensional, volume-averaged distribution of lithium ions and
potential in the solid and solution phases across a positive electrode,
separator, and negative electrode cell sandwich. In each of the porous Incorporating physical knowledge into the validation of many
electrodes, lithium ion intercalation is assumed to be governed by computations.—The dataset used in this work currently contains
Fickian diffusion into spherical particles coupled to faradaic charge 38,800 impedance spectra each containing 25 frequencies. As the
transfer via Butler-Volmer kinetics. The non-faradaic capacitive cur- number of computations grows to cover a wide range of parameter
rents assume a simple Helmholtz double-layer model. The full set and frequency space, deterministic methods for validating model ac-
of governing equations, boundary conditions, and parameter depen- curacy such as adaptive mesh refinement or manually verifying the
dencies are transformed into the frequency domain by assuming a number of node points in high gradient regions become impractical.
steady periodic solution form for all dependent variables driven by a Leveraging the physical understanding of the electrochemical system
single-frequency sinusoidal modulation of applied current. We apply provides a crucial insight into exploring the tradeoff between compu-
a Volterra series to capture the amplitude-dependence for the linear tational time and accuracy.
and weakly nonlinear higher harmonics that result from the single- As an example, in any porous electrode, the competing effects of
frequency input current perturbation. Further information showing the solid- and solution-phase conductivities and reaction kinetics de-
the model equations used for EIS spectra here are presented in the termine the distribution of current density within the system.40 That
Appendix AI of our prior work.37 is, for parameter sets with fast kinetics and low effective conduc-
The generic Coefficient Form PDE physics module in COMSOL tivities, the zone over which the lithium flux occurs is small, and
v4.4 was used here for all computations. As noted in earlier work,37 high gradients form. The location of the reaction zone is driven by
the highest density of nodes were placed at the interfaces between the ratio of solid- and solution-phase conductivities. Consequently,
the solid and solution phases and at the electrode/separator interfaces, for the lithium-ion battery system, where electrode conductivities are
based on results from preliminary analysis. For all results presented typically higher than electrolyte conductivity, the region of significant
here, the nodes are distributed in a geometric sequence with the node lithium flux density in the porous electrodes tends to be shifted toward
spacing at the electrode/separator interface 25X smaller than at the the electrode/separator interfaces. Similar considerations occur in the
center of the separator or at the current collector interface. All 38,800 diffusive flux of the oscillating lithium ion concentrations within the
computed spectra in our EIS library had the same mesh with 250 nodes solid electrode particles. The penetration depth is governed by Fickian
across the cell sandwich and 8000 elements in the particles. Because mass transport in the system,
of the size of the computed EIS library, we did not use mesh refinement
to validate that every individual spectrum was converged, as is normal 
Ds
when a small number of simulations (compared to 38,800) are used. δD ≈ [1]
We discuss our strategy for validation of the computational results in ω
§2.3. For a given set of physicochemical and geometric parameters
and single input perturbation frequency, computing the impedance where Ds is the solid-phase diffusivity and ω is the perturbation
response took approximately 4 seconds on a Dell Precision T1500 frequency. The boundary layer thickness decreases for lower mass
with an Intel Core i7 CPU @ 2.80 GHz and 8GB RAM using a transport coefficients and higher frequencies. To account for these
Windows 7 Professional 64-bit operating system. Parameters were regions of high gradients in the concentrations, potentials, and current
updated and the results were saved to disk using the LiveLink for densities in the system, a high number of node points at the interface
MATLAB version R2013b. The 38,800 member library of EIS spectra between the solid and solution phases and the electrode/separator
(with each spectrum having 25 frequencies) required approximately 2 interfaces were used.
CPU-months to compute, though much of it could be done in parallel. To establish confidence in the large EIS dataset we have produced,
we explore the most challenging computational limit (high frequen-
Creating a large dataset of simulated spectra.—The P2D cies) and compare it with the easily calculable analytic limits of infinite
impedance model as implemented here is parameterized with 26 frequency. In the limit of infinite frequency, the only contributor to
physicochemical and geometric coefficients. To adequately capture the impedance response is the combined ohmic drop across each of
the wide variety of impedance spectra one might encounter for the the electrodes and separator. The internal resistance of the cell has the
 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018) A299

analytical solution, ranked from lowest to highest error to determine the top matches for
    an experimental spectrum.
l pos lsep
R,predicted = +
σe f f, pos + κe f f, pos κe f f,sep
  Results
lneg
+ [2] Flexibility of physics-based modeling.—Due to the large ranges
σe f f,neg + κe f f,neg
over which the physicochemical and geometric parameters were sam-
where li is the thickness of the region, and σe f f,i and κe f f,i are pled, the resulting dataset of simulated impedance responses contains
the effective solid- and solution-phase conductivities (given by spectra of a wide variety of shapes in the Nyquist diagram. Figure 1
σ(1 − i −  f,i ) Br ugg and κi Br ugg where i and  f, i are the void- and shows some of the variation among the 38,800 computed spectra
filler-fractions and Br ugg is a Bruggeman-type tortuosity factor, re- demonstrating the flexibility of the P2D model to capture distinct
spectively). Eq. 2 sets a theoretical lower bound for the computed impedance responses. Figures 1a and 1b show the Nyquist represen-
electrochemical impedance for any set of parameters. Comparisons of tation of the simulated spectra with one overlapping (Fig. 1a) or two
high frequency computed solutions to Eq. 2 for the same parameter separable (Fig. 1b) kinetic arcs. The arcs are related to the interfa-
set is one way we will assess the numerical accuracy of our large cial charge transfer resistance and double-layer capacitance at the two
computed dataset over the wide parameter range used. electrodes. When the two kinetic impedance responses are similar in
magnitude, but differ in characteristic time constants, separate arcs
Fitting computational spectra to experimental data.—To directly (or a flat long arc) are visible. When either a single electrode has
compare a computed EIS spectrum to an experimental EIS spectrum extremely facile kinetics (low charge transfer resistance) or the elec-
from a battery with unknown electrode area requires a self-consistent trodes have similar kinetic time constants, a single arc defines the
method to determine the superficial electrode area, Amsup , for every mid-frequency impedance response. Consequently, a wide range of
computed spectrum m (here, our computed library has 1 ≤ m ≤ high- to mid-frequency responses are seen in the simulated dataset
38,800). Superficial area is needed because the computations provide without the requirement on introducing unphysical or nonideal circuit
area normalized impedances, with units in  − m2 , whereas EIS elements like the constant phase element. Similarly, despite the typical
experimental measurements have units in . For self-consistency, the treatment of fitting a single Warburg element, the low-frequency re-
area used to normalize EIS simulation results must produce a cell sponse also encompasses a diverse range of features. Figure 1c shows
capacity that matches the experimentally determined capacity of the an impedance spectrum with a small low-frequency response while
battery under test, Cdata (in Ah), which is a universally measured Figure 1d shows an impedance spectra with a comparatively large con-
or reported value. Thus, for every computed spectrum m, the basic tribution in the same frequency range. In the low-frequency regime,
procedure is to calculate each electrode’s capacity, C mpos and Cneg
m
, solid- and solution-phase mass transport as well as thermodynamics
according to (derivatives in the open circuit potential) interact to generate distinct
  variations in the impedance responses. As an example, Figure 1e and
Cim = Amsup lim 1 − im − mf,i Vi [3] Figure 1f show the effect of an impedance spectrum with a relatively
where the subscript index i is either pos or neg, Amsup is the unknown flat low-frequency feature and a steeply sloped feature, respectively.
superficial area (in m2 ), and Vi is the volumetric capacity (in Ah/m3 ) It should be noted that there is no need for the spectra in Figure 1 to
of the material in each electrode. The self-consistent superficial area be scaled by superficial area or shifted by inclusion of uncompensated
is the only unknown when the capacity-limiting electrode from each contact resistance since we are not comparing to experimental results,
computed spectrum m is equated to the experimental capacity data by and therefore, the figure units are  − m2 .
 
Cdata = min C mpos , Cneg
m
. [4] Assessing the quality of the simulation library.—Preliminary
One last factor often found in experimental data, but ab- studies were carried out to determine a meshing strategy that worked
m
sent from computed spectra, is contact resistances (Rcontact in for a wide range of parameters and frequencies. The goal of the pre-
). Uncompensated contact resistance leads to a shift along liminary work was to balance the number of mesh nodes and their
the real axis that will produce poor or erroneous fits to exper- spatial distribution, with the computational time required to compute
iments. Thus, we use contact resistance as a free parameter a converged solution. Given the wide range of parameters used, and
to minimize the residual error E m between the experimental the 38,800 unique parameter combinations explored at 25 frequen-
data and an area-scaled spectrum from the computed EIS dataset, cies (970,000 computations of the governing equations and bound-
ary conditions), it was not easy to evaluate the accuracy of every


  2   2
1    Z m ω j  
N
Z m ω j  
Em = + Rcontact
m
− Z data ωj + − Z data ω j [5]
N j Amsup m
Asup

m
where Rcontact is the error-minimizing contact resistance for spectrum computation. Nonetheless, we had a strategy to ensure the numerical
 
m, Z m (ω j ) and Z m (ω j ) are the real and imaginary components of approach and meshing was appropriate. In particular, we explored the
the computed impedance values for spectrum m at frequencies ω j , hardest-to-converge limit of the governing equations, high frequen-
 
Z data (ω j ) and Z data (ω j ) are the experimentally measured real and cies, by comparing the simulated high-frequency data to the analytical
imaginary impedance data at frequencies ω j , and N is the total number high-frequency limit given by Eq. 2.
of frequencies being fit. No weighting of the data with frequency is Figure 2 shows a comparison of the simulated high-frequency real
currently implemented. impedance at 105 Hz and the analytically predicted value of the ohmic
m
Because Rcontact is the only fit parameter to minimize error, it can resistance. An initial indicator of computational quality is seen imme-
be aphysical. We eliminate any spectrum from consideration if the diately; Eq. 2 is the theoretical lower bound for the real impedance, and
m
best-fitting Rcontact is negative and appreciable (here appreciable is we see that all simulations lie on or above the diagonal line, with none
taken to mean a magnitude greater than 10% of the high frequency below. There are two possible reasons for the simulated impedance at
limit). The residual error E m for all simulated spectra can then be 105 Hz to be above the analytically-predicted ohmic (real) resistance:
A300 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018)

Figure 1. Nyquist plots for six of the 38,800 simulated

impedance spectra. The spectra shown represent a wide range
of impedance responses across the input parameter space with
(a) small and (b) large separation between the kinetic arcs, (c)
small and (d) large low frequency Warburg-like “tails”, and
(e) flat and (f) steep low frequency responses. All spectra are
shown with 25 logarithmically spaced frequencies from 105
Hz to 10−3 Hz.

(i) the simulated frequency (105 Hz) is insufficient to reach a purely error for all spectra with less than a 5.7◦ phase angle at 105 Hz (68%
ohmic response for the set of parameters or (ii) numerical errors as- of the spectra) is 5.0%. In short, the closer a point at 105 Hz is to a
sociated with inadequate meshing. Understanding the origin of the purely real number, the more accurately Eq. 2 predicts the value. This
upward deviations we see in Fig. 2 at low cell resistances is aided by may seem obvious, but there is no reason to believe that inadequate
coloring the markers to indicate the computed phase of the calculated mesh refinement would produce this systematic behavior across such
impedance at 105 Hz for all 38,800 simulations. We see a consistent a diverse set of parameters.
trend of higher phase angles resulting in greater deviation from the To further test the idea that the meshing and computations are
diagonal line which is consistent with the deviation being associated adequate, and deviations arise predominantly from the selection of
with 105 Hz as too low of a frequency to reach the real axis. We can 105 Hz as the highest frequency, we further evaluated some of the
explore this effect statistically. The average deviation from the pre- largest deviating points. The inset in Fig. 2 shows the decrease in
dicted value for all 38,800 high frequency points is 12.9%, while the error resulting from increasing the highest simulated frequency from
100 kHz to 100 MHz, while keeping the grid mesh fixed. We see
that all three points move toward the diagonal line, further validating
that it is primarily the frequency range used, for certain parameter
combinations, not numerical error.
Based on these high frequency and other tests, we have confidence
that the systematic deviations seen in Figure 2 are not primarily due
to numerical error. Moreover, given that we primarily tested the most
difficult-to-compute frequency limit of the library, we are confident
that meshing-produced numerical errors are at least an order of mag-
nitude below the 5.0% systematic deviation seen for the low phase
angle points in Fig. 2, for all frequencies. Of course, adaptive methods
for increasing the density of node points within the high gradient re-
gions would lower the numerical error further, albeit at the significant
cost of increased computational time, but there appears little reason
to pursue that based on these results.

Using the ImpedanceAnalyzer: An example physics-based fit of

a LiCoO2 /graphite cell spectra.—The input panel for the web-based,
ImpedanceAnalyzer tool is shown in Figure 3. To fit an experimental
spectrum, a user would browse for a file on their computer con-
Figure 2. Comparison between the 38,800 simulated data points for the high- taining a comma delimited file where the first column contains fre-
est frequency used here, Z  (105 Hz), and the analytic ohmic resistance for the quency, the second contains the real impedance, and the third contains
same parameters, R, pr edicted . The marker color indicates the simulated phase the imaginary impedance. The type of analysis to be performed is
angle at 105 Hz. Inset shows that some combinations of parameters require a then selected and any additional input (battery capacity for scaling
much higher simulated frequency to achieve a purely ohmic response. according to Eq. 4, for example) is entered before clicking the “Go”
 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018) A301

Figure 4. Nyquist plots of the (a) raw experimental spectra, (b) interpolated
fit points, and (c) best matching simulated spectra (m = 6230) from the
38,800 computed datasets. The average error between the simulated spectra
and experimental data in (c) is 1.73% with a contact resistance of 4.2 m and
Figure 3. Input panel of the ImpedanceAnalyzer. Users would select a file to superficial area of 300.32 cm2 .
upload or an example dataset as well as analysis to perform before selecting
“Go”. Analysis menu items in gray are features we encourage the open source
community to add. matching” solutions can also be explored. Quickly visualizing the
many spectra that are close to matching the experimental spectrum
enables an experimentalist to begin piecing together a better under-
button to transmit the data to the server for the spectrum matching and standing of the interacting physicochemical processes in the complex
ranking process described above. electrochemical system. Figure 5 shows the residual errors calculated
To demonstrate the process, an example experimental impedance using Eq. 5 for the top 50 simulated spectra as a function of the spec-
spectrum is taken from the University of Maryland’s Center for Ad- tra’s rank. Multiple results with the same residual error occur when
vanced Life Cycle Engineering (CALCE) Battery Data Archive.41 The the only change in the input parameter set is a parameter that the lin-
spectrum shown here comes from the initialization impedance mea- ear impedance response is insensitive to (such as the charge transfer
surements on a 1500mAh LiCoO2 /graphite battery cell (Cell 41 in the coefficients, αa and αc , for example, where only their sum appears
PLN Initialization Dataset).42 The impedance response of the pouch in the governing equations and boundary conditions). The steepness
cell was measured from 1.64 kHz to 12.5 mHz. The Nyquist repre- of the drop-off at low rankings (best fitting spectra) indicates that
sentation of the raw experimental data is shown in Figure 4a. The there is still potentially significant benefit to increasing the number of
spectrum consists of a depressed semicircle in the frequency range simulations in the dataset.
associated with interfacial processes and a low-frequency tail in the
region associated with mass transport and thermodynamic processes.
As a part of the spectrum ranking process, the experimentally sam-
pled frequencies are quadratically interpolated to match the simulated
frequencies (Figure 4b) and then the process described above deter-
mines the simulated spectra from the dataset with the lowest residual
error (Figure 4c). The resulting closest match for this example spectra
and the P2D dataset described above has a run index, m, of 6230, a
superficial area, Asup , of 300.3 cm2 , a contact resistance, Rcontact , of
4.21 m, and an average residual error, E, of 1.73%. In this case,
the positive electrode provides the limiting capacity. The parameters
which generate the closest match are returned to the user in the web
tool and are shown in Table AII in the Appendix. It should be noted
that the experimental dataset used here has the data truncated at a mod-
erate frequency and, thus, the spectrum contains negligible influence
from the inductance often found experimentally at high frequencies.
For datasets that contain higher frequencies, there may need to be a
simple inductive element added to the contact resistance parameter in
Eq. 5 to capture these experimental artifacts.
Figure 5. Plot of the top 50 ranked residual average errors between the exper-
Exploring the remaining solutions.—One of the benefits of the imental data set and the 38,800 simulated spectra. The letters denote (a) 3rd ,
dataset-based approach taken in this work is that the nearby “nearly- (b) 7th , and (c) 10th ranked fitting errors corresponding to the spectra in Fig. 6.
A302 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018)

Figure 7. Nyquist plot comparing the experimental data with a two parameter
locally optimized P2D simulated spectrum. A 20% increase in the double layer
capacitance, Cdl,neg , and a 50% decrease in the solid diffusion coefficient,
Ds,neg , resulted in a spectrum with a 1.25% error (33% lower than the best
matching spectra in the 38,800 member dataset).

match. Interestingly, due to the symmetry of the P2D model, it is

possible to switch the labels of pos and neg parameters and compute
an identical spectrum. Additionally, the estimated superficial area for
each of these nearly matching spectra vary more than 100 cm2 . Conse-
quently, including additional information about the battery (including
any known physicochemical or geometric ranges of parameters) will
be useful for improving the matching process by filtering the dataset.
Adding this feature is currently under development.

Matched spectra as an initial guess for local optimization.—Sam-

pling across a large range of physical parameters enables a fit such as
that shown in Figure 4c, but, because we used parameters associated
Figure 6. Nyquist plots of the (a) 3rd (m = 34560), (b) 7th (m = 4932), and with a wide range of different battery chemistries, states-of-charge,
(c) 10th (m = 32014) best matching spectra in the dataset of 38,800 simulated etc., the 38,800 member EIS library is not a dense covering of the
spectra. parameter space. Moreover, for any finite number of simulations, the
global nature of the sampling process makes it likely that the result-
To further demonstrate the different variations by which the sim- ing match is not a best-fit to the experimental spectrum. However,
ulated spectra fail to match the experimental data and how we can if there is low experimental noise, the selected spectrum makes a
learn from them, the three spectra in Figure 6 show the starred 3rd , good starting point for further local optimization of the physics-based
7th , and 10th closest matching spectra in the dataset from Figure 5. model. In the case of the spectrum shown in Figure 4c, the largest
For example, while Figure 6a (m = 34560) matches well in the low- contributions to the remaining 1.73% error are in the low-frequency
frequency region, it fails at matching in the high-frequency regime. tail as well as the kinetic arc at high frequencies. Using conventional
Conversely, the spectrum in Figure 6b (m = 4932) qualitatively understanding of impedance spectra, the width of the high frequency
matches, but has the timescales for kinetics shifted, while the spec- arc can be attributed to the charge transfer resistance associated with
tra shown in Figure 6c (m = 32014) has a qualitatively similar the intercalation kinetics in both electrodes, while the time constants
response in the high-frequency regime, but fails at low frequencies. of the arcs (dictating their separation and shape) are also influenced
Interactively exploring these nearby solutions is made easy in the by the double-layer capacitances. Thus, to change the shape of the
ImpedanceAnalyzer via an “Explore P2D” modal where a user can kinetic arcs without altering the total width, the double-layer capac-
mouse over the residual points and interactively see the corresponding itances can be varied. Additionally, because the time constants for
spectrum and parameter values. Clicking on a residual point allows kinetics and mass transport are separated by more than an order of
the comparison of multiple spectra and a download of the selected magnitude, changes in either parameter group only affect the high-
parameters is made available. The three parameter sets for the spec- frequency arc and low-frequency tail, respectively. Figure 7 shows
tra in Fig. 6 are presented in Table AIII. We see that the qualitative the consequence of a simple iterative reduction of residual error by
improvement in the low-frequency response of the 3rd best spectra varying the double-layer capacitances as well as the solid-phase dif-
is likely due to the decreasing solid-phase diffusion coefficient in fusion coefficients and open circuit potential derivatives. For a 20%
the negative electrode, Ds,neg , while further decreasing the diffusion increase in double-layer capacitance, Cdl,neg , and a 50% decrease in
coefficients many more orders of magnitude leads to the very poor the negative electrode solid diffusion coefficient, Ds,neg , the residual
match at low-frequencies for the 10th best spectrum. Additionally, error decreased to 1.25%. Directly incorporating sophisticated local
decreasing the magnitudes of the open circuit potential derivatives optimization in the vicinity of several highly ranked solutions found
leads to a smaller low-frequency tail between the 3rd and 7th best by global sampling is a functionality currently being explored for the
matching spectra. For the mid-frequency response, the lower negative ImpedanceAnalyzer.
exchange current density, i 0,neg , leads to similar time constants be-
tween electrodes resulting in a single, narrower arc for the 3rd ranked
Implications and Concluding Remarks
spectrum with respect to the best match. Decreased double-layer ca-
pacitances (increasing the characteristic frequency of both electrodes) The physics-based nature of the P2D and other electrochemical
broadens the mid-frequency response of the 7th ranked spectrum into models enables a wide set of physically relevant and interpretable
a single, flatter kinetic arc, while increasing the positive electrode impedance responses to be captured by simply varying the combi-
capacitance (lowering the characteristic frequency) and decreasing nations of physicochemical parameter inputs. With a physics-based
the negative electrode capacitance (increasing the characteristic fre- approach, it is transparent what phenomena are included in the model
quency) removes the separation between the two electrode arcs in the and what assumptions underpin the model. In contrast, equivalent cir-
10th best fit. cuits are reduced order models of (normally) complex physical and
The limiting electrode (i.e. the electrode which has the computed chemical processes, sometimes leading to ambiguous or incorrect in-
area of 1500 mAh according to Eq. 4) in each of the simulations terpretation of results.28 While it is normally possible to fit some form
shown in Figure 6 is the positive electrode except for the 10th best of an equivalent circuit to an experimental spectrum, linking the fit to
 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018) A303

the underlying phenomena for anything but the simplest electrochem- Table AI. (Continued.)
ical systems remains a fraught exercise.
In the work described here, the P2D battery model was used to Physicochemical Parameters
demonstrate the dataset-based approach for parameter estimation. As Parameter Lower Limit Upper Limit Units
noted in the introduction, the high citation rate of the P2D model makes
it the de facto “standard” continuum model for capturing the dynamics c0 100 5000 mol/m3
of the lithium-ion system. Moreover, significant research efforts have D 10−11 10−9 m2 /s
focused on adding additional physics to the system making it a good D s,n 10−19 10−10 m2 /s
candidate for demonstration. That said, the open-source, dataset-based D s, p 10−17 10−10 m2 /s
dU n
approach described here enables an opportunity to start statistically d cs −100 0 V · cm3 /mol
dU p
comparing electrochemical models as different physical processes are d cs −100 0 V · cm3 /mol
included or removed. Combining the sampling scheme with methods i 0,n 1 20 A/m2
for performing a global, variance-based sensitivity analysis43 of the i 0, p 1 200 A/m2
simulated dataset can provide information on the parameters most re- t 0+ 0.2 0.99 –
sponsible for variations in the impedance spectra at a given frequency. α a,n 0.2 0.8 –
It has been shown that parameter identifiability is an important con- α a, p 0.2 0.8 –
sideration in fitting electrochemical models to experimental data.44 κ 0.1 8 S/m
Comparing the parameter sensitivity as more complex physical inter- σn 30 4000 S/m
actions are sequentially added to models can provide statistical insight σp 4 30 S/m
into the tradeoff between a more descriptive physical representation Geometric Parameters
of the system and potentially worse parameter identifiability as addi- ln 15 1200 μm
tional parameters are added. Additionally, comparing an experimental ls 8 40 μm
spectrum to the nearest matching spectra within a single dataset can lp 15 1200 μm
enable a bootstrap approach to establishing confidence intervals in the Rn 0.25 20 μm
estimated parameters. Rp 1 15 μm
Several additional benefits arise from the dataset-based approach  f ,n 0.0 0.12 –
 f,p 0.0 0.12 –
described here. The computational time is spent upfront during the
n 0.2 0.5 –
generation of the dataset, significantly reducing the time an experi-
s 0.35 0.8 –
mentalist waits to find a best matching spectrum. The time needed to p 0.2 0.5 –
identify the best matching fit to experimental data is unchanged by
the complexity of the model. Only the size of the library affects the
estimation time. Of course, a more complex model will take longer to
compute the library of spectra; however, separating the model com-
plexity from the time required to estimate the parameters becomes Table AII. Physicochemical and geometric parameter for
increasingly important as the quest for deeper insight into complex matching spectra.
electrochemical systems grows. Furthermore, the open-source nature
of the platform enables the inclusion of additional electrochemical Fit Parameters
data such as higher order harmonic responses,37 spectra at multiple Parameter Value [Units]
states (depths-of-discharge or temperatures), as well as (dis)charge
curves or cyclic voltammetry. Asup 300.3 [cm2 ]
The power of open-source software like this lies in the opportunity R cont act 4.2 [m]
for subsequent contributors to add features, additional simulated or Physicochemical Parameters
experimental data sets, etc., thereby evolving the tool to a more useful C dl,n 79.8 [μF/cm2 ]
form. To that end, the ImpedanceAnalyzer is simply the start of a C dl, p 86.5 [μF/cm2 ]
platform to which others are encouraged to contribute. The code is c0 1485.3 [mol/m3 ]
openly available on GitHub and users of the software can cite the D 7.68E-10 [m2 /s]
version released at the time of this manuscript.45 Other open software D s,n 9.09E-12 [m2 /s]
and open data products being actively built by the electrochemical D s, p 4.34E-16 [m2 /s]
data science community are also available for those interested in dU n
−40.5 [V · cm3 /mol]
d cs
collaborative development.46 dU p
−9.60 [V · cm3 /mol]
d cs
i 0,n 19.42 [A/m2 ]
i 0, p 1.56 [A/m2 ]
Acknowledgments t 0+ 0.45 [–]
Partial support of this research is provided by the U.S. Depart- α a,n 0.58 [–]
ment of Education grant P200A120023, an NSF IGERT grant DGE- α a, p 0.41 [–]
1258485, and the Boeing Endowment for Excellence. Data and the κ 1.42 [S/m]
analysis code for this paper can be found online.47 σn 1986 [S/m]
σp 19.6 [S/m]
Geometric Parameters
Appendix ln 319.9 [μm]
ls 35.9 [μm]
lp 159.1 [μm]
Table AI. Physicochemical and geometric parameter ranges. Rn 19.3 [μm]
Rp 1.49 [μm]
Physicochemical Parameters  f ,n 0.062 [–]
Parameter Lower Limit Upper Limit Units  f,p 0.075 [–]
n 0.366 [–]
C dl,n 1 100 μF/cm2 s 0.680 [–]
C dl, p 1 100 μF/cm2 p 0.354 [–]
A304 Journal of The Electrochemical Society, 165 (2) A297-A304 (2018)

Table AIII. Physicochemical and geometric parameter for nearly 5. Q. Wang, J. -E. Moser, and M. Grätzel, J. Phys. Chem. B, 109, 14945 (2005).
6. D. D. Macdonald, S. Real, S. I. Smedley, and M. Urquidi-Macdonald, J. Electrochem.
matching spectra shown in Fig. 6.
Soc., 135, 2410 (1988).
7. D. D. Macdonald, Electrochimica Acta, 51, 1376 (2006).
Fit Parameters 8. M. D. Levi and D. Aurbach, J. Phys. Chem. B, 101, 4630 (1997).
Rank: 3 Rank: 7 Rank: 10 9. E. Warburg, Ann. Phys. Chem., 303, 493 (1899).
Parameter Units Run: 34560 Run: 4932 Run: 32014 10. M. D. Levi and D. Aurbach, J. Phys. Chem. B, 101, 4641 (1997).
11. R. De Levie, in Advances in Electrochemistry and Electrochemical Engineering Vol.
Asup cm2 313.9 268.7 218.4 6, P. Delahay Editor, p. 329, Interscience, New York (1967).
12. M. Musiani, M. Orazem, B. Tribollet, and V. Vivier, Electrochimica Acta, 56, 8014
R cont act m 10.1 4.9 8.0 (2011).
Physicochemical Parameters 13. N. Fouquet, C. Doulet, C. Nouillant, G. Dauphin-Tanguy, and B. Ould-Bouamama,
C dl,n μF/cm2 31.0 57.2 32.3 J. Power Sources, 159, 905 (2006).
14. R. Onanena, L. Oukhellou, D. Candusso, F. Harel, D. Hissel, and P. Aknin, Int. J.
C dl, p μF/cm2 30.7 47.3 97.3 Hydrog. Energy, 36, 1730 (2011).
c0 mol/m3 4635 4921 986.5 15. P. Singh, R. Vinjamuri, X. Wang, and D. Reisner, Electrochimica Acta, 51, 1673
D m2 /s 8.547E-10 7.588E-10 1.943E-11 (2006).
D s,n m2 /s 2.465E-13 9.989E-18 2.793E-19 16. B. Saha, K. Goebel, S. Poll, and J. Christophersen, IEEE Trans. Instrum. Meas., 58,
291 (2009).
D s, p m2 /s 3.214E-15 8.104E-15 2.205E-17
dU n 17. G. Cui, H. Liu, G. Wu, J. Zhao, S. Song, and P. K. Shen, J. Phys. Chem. C, 112, 4601
d cs V · cm3 /mol −0.1593 −0.1937 −2.976 (2008).
dU p
d cs V · cm3 /mol −43.03 −26.77 −0.02543 18. J. Segalini, B. Daffos, P. L. Taberna, Y. Gogotsi, and P. Simon, Electrochimica Acta,
55, 7489 (2010).
i 0,n A/m2 2.619 19.68 6.163
19. R. J. Sheppard, B. P. Jordan, and E. H. Grant, J. Phys. Appl. Phys., 3, 1759 (1970).
i 0, p A/m2 1.616 1.934 1.937 20. B. A. Boukamp, Solid State Ion., 20, 31 (1986).
t 0+ – 0.984 0.483 0.942 21. J. E. B. Randles, Discuss. Faraday Soc., 1, 11 (1947).
α a,n – 0.524 0.528 0.348 22. S. Fletcher, J. Electrochem. Soc., 141, 1823 (1994).
α a, p – 0.632 0.269 0.264 23. B. Tribollet, J. Electrochem. Soc., 130, 2016 (1983).
24. D. T. Schwartz, P. Stroeve, and B. G. Higgins, J. Electrochem. Soc., 136, 1755 (1989).
κ S/m 5.726 7.024 6.544
25. Q. Guo and R. E. White, J. Electrochem. Soc., 151, E133 (2004).
σn S/m 356.1 173.7 2484 26. J. R. Wilson, D. T. Schwartz, and S. B. Adler, Electrochimica Acta, 51, 1389
σp S/m 10.63 5.638 22.71 (2006).
Geometric Parameters 27. J. P. Meyers, M. Doyle, R. M. Darling, and J. Newman, J. Electrochem. Soc., 147,
2930 (2000).
ln μm 422.6 262.1 286.7 28. M. Doyle, J. P. Meyers, and J. Newman, J. Electrochem. Soc., 147, 99 (2000).
ls μm 17.85 8.422 28.35 29. S. Devan, V. R. Subramanian, and R. E. White, J. Electrochem. Soc., 151, A905
lp μm 150.1 142.9 336.4 (2004).
Rn μm 2.447 18.45 6.418 30. D. Dees, E. Gunen, D. Abraham, A. Jansen, and J. Prakash, J. Electrochem. Soc.,
Rp μm 6.117 6.23 2.384 152, A1409 (2005).
31. B. R. Hough, D. T. Schwartz, and J. Pfaendtner, Ind. Eng. Chem. Res., 55, 9147
 f ,n – 0.094 0.041 0.069 (2016).
 f,p – 0.011 0.028 0.033 32. M. Doyle, T. F. Fuller, and J. Newman, J. Electrochem. Soc., 140, 1526 (1993).
n – 0.234 0.331 0.332 33. T. F. Fuller, M. Doyle, and J. Newman, J. Electrochem. Soc., 141, 1 (1994).
s – 0.479 0.757 0.494 34. D. P. Abraham, S. Kawauchi, and D. W. Dees, Electrochimica Acta, 53, 2121 (2008).
p – 0.411 0.262 0.271 35. http://jes.ecsdl.org/reports/most-cited.
36. I. M. Sobol’, USSR Comput. Math. Math. Phys., 7, 86 (1967).
37. M. D. Murbach and D. T. Schwartz, J. Electrochem. Soc., 164, E3311 (2017).
38. J. Herman and W. Usher, J. Open Source Softw., 2, 97 (2017).
39. B. Hough, C. Fu, and S. Paliwal, savvy (Version v2.0), Zenodo, (2016).
ORCID 40. J. S. Newman and K. E. Thomas-Alyea, Electrochemical systems, 3rd ed., J. Wiley,
Hoboken, N.J, (2004).
Matthew D. Murbach https://orcid.org/0000-0002-6583-5995 41. S. Saxena, C. Hendricks, and M. Pecht, J. Power Sources, 327, 394 (2016).
42. http://www.calce.umd.edu/batteries/data/.
43. A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana,
References and S. Tarantola, Global sensitivity analysis: the primer, Wiley, New York, NY,
(2008).
1. E. Barsoukov and J. R. Macdonald, Eds., Impedance spectroscopy: theory, ex- 44. A. M. Bizeray, J. -H. Kim, S. R. Duncan, and D. A. Howey, ArXiv Prepr.
periment, and applications, 2nd ed., p. 595, Wiley-Interscience, Hoboken, N.J, ArXiv170202471 (2017).
(2005). 45. M. D. Murbach and D. T. Schwartz, ImpedanceAnalyzer (Version v0.9.0-beta), Zen-
2. M. E. Orazem and B. Tribollet, Electrochemical impedance spectroscopy, p. 523, odo, (2017).
Wiley, Hoboken, N.J, (2008). 46. M. D. Murbach, C. Hendricks, W. A. Beckner, J. Chen, M. Yess, D. T. Schwartz,
3. V. F. Lvovich, Impedance spectroscopy: applications to electrochemical and dielec- D. A. C. Beck, C. Karuppaiah, P. Stanley, S. Ramanathan, S. C. DeCaluwe, L. Sprowl,
tric phenomena, p. 353, Wiley, Hoboken, N.J, (2012). L. Tsui, Y. Liu, T. Das, and M. Dixit, ECS Data Science Hack Day, (2017).
4. F. Fabregat-Santiago, J. Bisquert, G. Garcia-Belmonte, G. Boschloo, and A. Hagfeldt, 47. M. D. Murbach and D. T. Schwartz, Data and code for manuscript (Version v1.0),
Sol. Energy Mater. Sol. Cells, 87, 117 (2005). Zenodo, (2017).