We demonstrate that failure of time-invariance assumption in the modeling of electrochemical systems by equivalent circuits can lead to the formation of low frequency "inductive loops" that manifest themselves as positive imaginary parts of the impedance function. Assuming that the properties of the equivalent circuits change slowly in time we perform an asymptotic analysis and obtain a new integral representation of the impedance function that reduces to the standard one at high frequencies, while exhibiting inductive loops at low frequencies.
Introduction
Electrochemical impedance spectroscopy (EIS) is an indispensable tool to describe complex electrochemical systems in a unified and graphical way. In this approach any system is described by a single complex-valued impedance function Z(ω). The basic theory of EIS [3, 4] [19] says that this function possesses special analytic properties that are most concisely expressed by the representation
Here σ is a positive Borel-regular measure on [0,+∞). It is often convenient to approximate such measures by linear combinations of delta-functions, resulting in rational approximations of Z(ω)[1]. Such approximations can be interpreted as impedances of electrical circuits made of resistors and capacitors only. They are called the equivalent circuit models (ECM). One easily verifiable feature of the representation (1) is the negativity of the imaginary part of the impedance function.
The experimentally observed arcs in a Nyquist plot of the impedance with positive imaginary part in the low frequency part of the spectrum received the unfortunate moniker of an “inductive loop” [8] [12, 23, 21, 10]. It is generally agreed that magnetic effects play no significant role in the functioning of most electrochemical systems and devices, and even when they do, the inductive effects manifest themselves at the high end of the frequency spectrum [7, 6] [16]. The explanation of inductive loops in the current literature [6] [16] by means of ECMs with negative resistances and capacitances creates more problems than it solves. If one permits negative resistances and capacitances, then arbitrary “experimental data” can be matched by such an impedance function with any degree of precision due to the Riesz theorem (see e.g. [17, 14, 15]). This is not surprising, since it is the passivity of the system [22, 9, 20], violated by the the negative resistances and capacitances in ECMs, that are responsible for the possibility of stable reconstruction of the impedance function.
A far more reasonable explanation of the experimental appearance of low frequency inductive loops is the violation of the time-invariance assumption at time scales commensurate with the inverses of low frequencies. This explanation is semi-explicit in [21, 16, 18], where specific electrochemical processes altering the properties of the system on slow time scales are identified.
When the time-invariance assumption is violated the input-output behavior of the system can no longer be mathematically described by a well-defined impedance function at low frequencies. In this paper we will show that if one applies a particular impedance measurement procedure, delivering the correct impedance for any time-invariant system, the inductive loops could be observed in non time-invariant systems.
The goal of this paper is to extend the classical EIS/ECM approach to electrochemical systems with slowly “drifting” properties. In § 3 we demonstrate that a single Voigt element with drifting properties may exhibit an inductive loop. In § 5 we use asymptotic analysis to extend the EIS/ECM approach to non time-invariant systems with slowly varying properties. The resulting impedance representation formula, though more complicated, may still be used to reconstruct the impedance function values within the frequency band containing experimental measurements by performing the classical Kramers-Kronig analysis [2, 5] [11, 20] for high frequency data, while fitting the additional “drift” parameters using the inductive loop low frequency data.
The paper is organized as follows. In § 2 we discuss a measurement technique that gives a good approximation of the true impedance function for time-invariant systems. We then show in § 3 that the same measurement technique for non time-invariant Voigt elements can produce “inductive loop” data. Representing a general non time-invariant electrochemical system by an ECM with non time-invariant Voigt elements connected in series, gives us a mathematical model that can be analyzed. An asymptotic analysis applied to each non time-invariant Voigt element leads us in § 5 to a generalization of the EIS theory that reduces to classical formulas at high frequencies and is capable of modeling low frequency inductive loops.
Impedance measurements
Let us assume that the input signal is the current
I(t)=I0sin(ωt),t∈[0,T(ω)],T(ω)=ω2πn(ω),(2)
where n(ω)∈N should be as large as possible, so that the time T(ω) it takes to make a measurement is still acceptable to whoever makes the measurements. Of course, this is an issue only for low frequencies ω. Thus, we can measure the system response only for frequencies ω≥ωmin=2π/Tmax, where Tmax is the maximal admissible time for making low frequency measurements. Hence, in the low frequency regime we always choose n(ω)=1. At high frequencies we use a different measurement strategy. We choose a fixed measurement time T0 and set n(ω)=round(T0ω/2π). Hence, for all ω≥ωmin we define
n(ω)=max{1,round(2πT0ω)},ω≥ωmin.(3)
The Fourier transform of I(t) can be computed explicitly, but is an unwieldy expression, except at the frequency ω of the input sinusoid:
For linear, time-invariant (LTI) systems the output voltage U(t) must satisfy U(ω)=Z(ω)I(ω), where the complex factor Z(ω) is called the impedance. Theoretically, when I(t)=0 for t>T(ω), the output voltage is still non-zero. However, in most cases it decays exponentially fast and can therefore be neglected. In this case we have the approximation
U(ω)≈∫0T(ω)U(t)e−iωtdt.(5)
Since U(t) is measured, formula (5) can be used to estimate U(ω) and thus, we have the formula for estimating the impedance from the experimental data:
Z0(ω)=T(ω)2i∫0T(ω)U(t)e−iωtdt.(6)
If we apply formula (6) to an elementary Voigt circuit, consisting of a resistor R and a capacitor C connected in parallel, we will not get the correct answer
ZV(ω)=iθω+1R,θ=RC.(7)
Figure 1 shows that Z0 is not a very good approximation for ZV, especially in the intermediate frequency range. The remedy is to understand this discrepancy and then devise a way to correct for it. Using the explicit expression for Z0(ω) (which we don’t display here), we discover that for an elementary Voigt circuit we have
where ZV′(ω) denotes the derivative of ZV(ω) with respect to ω. Since relation (8) between the measured and the true impedance function of an elementary Voigt circuit is linear, it extends to all LTI systems.
We recall that in an LTI system the output (voltage U(t)) depends on the input (current I(t)) via
U(t)=ρ0I(t)+∫−∞tI(τ)K(t−τ)dτ,(9)
where the function K(s) is called a memory kernel.
Theorem 2.1.Suppose that the memory kernel decays exponentially:
Thus, we are lead to a simple method for estimating the true impedance by making measurements over a longer time T2=2πn2(ω)/ω>T1=2πn1(ω)/ω. Then, up to an exponentially small error we must have
We still want to require that at high frequencies the time it takes to make a an impedance measurement is fixed. Let us call it T0′, since we still want n1(ω) to be given by (3). In that case
Thus, for measuring very low frequencies we set n1(ω)=1 and n2(ω)=2. When ω=4π/T0 we have n1(ω)=2, in which case we must require that n2(ω)≥3. It follows that T0′≥3T0/2, and therefore, n2max(ω)>n1(ω). In practice, the experimenter can choose any integer n2(ω) satisfying n2min(ω)≤n2(ω)≤n2max(ω), once the constant T0′≥3T0/2 has been set.
The correction method (13) can be interpreted as a directive to ignore the transient response over the time interval [0,T1], effectively taking data during time interval [T1,T2]. Indeed, using formula (6), we have
One convenient choice is n1=n(ω) and n2=2n(ω), where n(ω) is given by (3), so that the length of the time interval over which the data is taken is still T(ω)=2πn(ω)/ω.
Theorem 2.2. Under assumptions of Theorem 2.1 we have the estimate
To see how formula (15) improves the evaluation of the impedance we take the same Voigt circuit with R=2kΩ, C=200μF, shown in Figure 1, and compare Zexp(ω) and Z(ω)=ZV(ω). The difference between the Nyquist plots of the two functions can no longer be visualized as in Figure 1, since it is less that 0.01%, according to numerics, when we choose n1=n(ω) and n2=2n(ω).
Image not extracted
Figure 1. Comparison of Z0(ω) and Z(ω) for an elementary Voigt circuit with R=2kΩ, C=200μF.
Non time-invariant elementary Voigt circuits
In this section we examine elementary Voigt circuits made of a resistor and a capacitor connected in parallel. Our main assumption is that the parameters R and C of the circuit do not stay constant, but slowly change in time, instead. We will show that if we apply the impedance measurement recipe from the previous section to such non time-invariant systems we may obtain impedance curves with inductive loops.
General theory
Let us start by considering an elementary Voigt circuit consisting of a resistor R=R(t) and a capacitor C=C(t) connected in parallel. In such a circuit we have the Ohm’s laws for each of the elements
IR(t)=R(t)U(t),IC(t)=C(t)U˙(t).
Combining this with the Kirchhoff’s law I(t)=IR(t)+IC(t) we obtain the constitutive relation (i.e. dependence of I(t) on U(t)) in the form
I(t)=R(t)U(t)+C(t)U˙(t).(17)
Since we use the current as the input we need to solve (17) for U(t):
U(t)=∫−∞tC(s)I(s)exp{−∫stθ(x)dx}ds,θ=RC.(18)
The quantity θ is called the relaxation time of the Voigt circuit. Formula (15) gives the experimentally measured “impedance” of a simple Voigt circuit with non time-invariant elements:
Let us now examine a particular model of the time dependence of R(t) and C(t). We assume that R and C undergo an “exponential drift” from R− (resp. C−) at t=−∞ to R+ (resp. C+) at t=+∞:
The evolution law for R(t) is governed by two parameters: the time scale τ and the “current position” a>0 that tells us how far along R(t) is on the way from R− to R+. It is in 1-1 correspondence with R(0). The value a=1 means R(0) is exactly half-way between R+ and R−. Exact same comments apply to C(t). It is easy to calculate explicitly
is a limiting case of our model (22) with R+=R∞, C+=C∞, and
a,b→0,R−,C−→∞,aR−→R0−R∞,bC−→C0−C∞.
In this limit, however, the restriction that R− and C− are positive are no longer required, since R(t) and C(t) are required to be positive only for t≥0. In this drift model we can still use (23) with
The experimentally measured impedance Zexp(ω) is then computed by means of formula (15). The numerical evaluation of Zexp(ω) presents several challenges, since parameter τ must be very large, relaxation times θ± are typically very small and ω can range from 10−6Hz to 106Hz, making some integrands highly oscillatory. These issues are addressed in Appendix B. Figure 2 shows an inductive loop in the Nyquist impedance plot of a non time-invariant Voigt element, with R(t) and C(t) given by (22), where R+=0.4kΩ, R−=2kΩ, a=10, C+=1.6mF, C−=0.8mF, b=6, τ=600s. The figure shows that at high frequencies the impedance is very well captured by the impedance of the time-invariant Voigt element with R=R(0), and C=C(0).
Slow parameter drift asymptotics
Assume that 1/τ=ϵ is a small parameter and that ω=ϵΩ, where the rescaled frequency Ω is assumed to be fixed. We need to find the asymptotics of Z(ϵΩ) as ϵ→0 and see if Im(Z(ϵΩ)) can indeed be positive. We now assume that R(t)=R0(ϵt) and C(t)=C0(ϵt), where R0(x) and C0(x) are strictly positive, bounded smooth functions with bounded first derivatives on R. Our goal is not only to obtain the asymptotics of Z(ω), defined by (6), as ϵ→0, but estimate the error between the actual impedance and its asymptotic approximation. We obtain for the elementary Voigt circuit, using (18)
where the order ϵ error O(ϵ) is uniform over the entire frequency spectrum.
Formula (28) shows that if the functions θ0(η) and R0(η) are constants, then the approximation becomes exact, as it reduces to the classical impedance of the elementary Voigt circuit. One more simplification is possible in the asymptotics ϵ→0. Specifically, we can replace θ0(η) by θ(0) in (28). Our final asymptotics is
For numerical evaluation of (29) we choose n1=n(ω), n2=2n(ω) and change variables of integration η=sn/Ω. Then, using the representation Im(z)=(z−zˉ)/2i we obtain the formula we use in the Matlab code to make Figure 3:
Image not extracted
Figure 3. The quality of the asymptotics of an inductive loop.
We can therefore, use a linear approximation R0(x)≈bR+mRx, when x∈[ϵT0,2ϵT0], where the parameters bR and mR are found from the least squares fit. In that case
2π1∫2π4πR0(Ωsn)e−2insds≈2ΩimR.
Figure 3 shows the Nyquist plot of Zϵ∗(ω) superimposed on the computed “measured” impedance corresponding to the model (22). The difference between the two graphs is not detectable at the scale of the entire Nyquist plot, shown in the left panel of the figure. The right panel shows the blown-up portion of the plot containing the inductive loop. The overall relative error of the asymptotic approximation (29) is about 0.1%. In the figure we used (22) with the same parameter values as in Figure 2.
Now, if we have a general electrochemical system exhibiting parameter drift, then the measured impedance can be represented as
where for each η∈R the function ω↦Z(ω,η) is a classical (instantaneous) impedance function, provided there is no leading capacitance in the ECM. In the presence of the leading capacitance, the term
Typically, one would choose n1(ω)=n(ω) and n2(ω)=2n(ω), where n(ω) is given by (3). At low frequencies n(ω)=1, while at high frequencies 2πn(ω)/ω can be replaced with a constant T0.
Conclusions and discussion
In this paper we have proposed an explanation of the inductive loop phenomena observed in experiments. The hypothesis is that the act of the impedance measurement of an electrochemical system might speed up slow processes, such as corrosion and charge diffusion, altering the properties of the system on time scales commensurate with the inverses of the frequencies at which the impedance is measured. The modeling and analysis of these processes can lead to specific drift laws that can be used to validate our general theory.
A natural question is whether our theory makes it apparent how the time-dependent nature of the system causes the occurrence of inductive loops. By way of the answer we refer to formula (30) and set ω=0, n(ω)=1. Then we obtain
Im(Z(ϵΩ))≈2π1∫2π4πR0(Ωs)sin(2s)ds(36)
While R0(u) is a strictly positive, smooth real function, the sign of the integral in (36) is the outcome of balancing positive and negative contributions of the oscillatory integrand. For example, if R0(u) is a decreasing function, then positive contributions will always be larger than the negative ones and inductive loops will be present. Conversely, if R0(u) is an increasing function, no inductive loops will be produced. If the slowly oscillating sinusoidal current input causes a corresponding in-phase oscillation of R0, the inductive loop effect could be several times stronger.
One open question, not directly addressed in the paper is capturing the time scale τ over which the system’s properties change. A crude estimate would be the inverse frequency at which the imaginary part of the impedance hits zero. For example, in the simulation corresponding to Figure 2 the time scale τ=600s, while 2π/ω0=610s, where Im(Z(ω0))=0. Our simulation of the time-dependent elementary Voigt circuit shows that the deviations from the classical model start at frequencies ω as large as two orders of magnitude over τ. They become vividly pronounced at frequencies on the order of 1/τ and persist over frequencies an order of magnitude lower.
Figure 2. An inductive loop in a non time-invariant Voigt element.
Finally, whether or not the inductive loop data can give additional information about the system beyond the time scale τ depends on whether the drift model (34) is acceptable. In this case one should be able to use the algorithm in [13] applied to the high frequency data to compute parameters Rj(0), θj(0) and N. The intermediate and low frequency data can then be used in a non-linear least squares fit to to estimate drift parameters μj and ΔRj. Future research into the inductive loops should address these questions.
Acknowledgments. This material is based upon work supported by the National Science Foundation under Grant No. DMS-2005538.
Mathematical proofs
Proof of Theorem 2.1 and 2.2
For I(t)=χ[0,T(ω)](t)sin(ωt) we have, making a change of variables s=t−τ,
U(t)=ρ0sin(ωt)+∫0tK(s)sin(ω(t−s))ds,0≤t≤T(ω).
We can then write
U(t)=ρ0sin(ωt)+∫0∞χ(s,+∞)(t)K(s)sin(ω(t−s))ds.
so that we can substitute this into (6) and switch the order of integration. We obtain
The function Eε0(ξ,η) is exponentially small when ξ is not very close to η. However, when ξ is very close to η we can replace C0(ξ) with C0(η). Hence we have
Hence, we obtain a uniform in the entire spectrum bound
∣Zε(ω)−Zε∗(ω)∣≤Mε.
for some constant M that depends on the parameter functions R0(s) and C0(s), and that scales like Rθ.
Numerical evaluation of the impedance in explicit drift models
The key to the effective numerical computation of Zexp(ω) for explicit drift models (22) and (26) is a non-dimensionalization of all quantities. We therefore begin by rescaling the variable of integration in the definition (19) of Z(ω), t=nt^/ω.
Z(ω)=I0πi∫2π4πU(ωnt^)e−int^dt^.
where we have chosen n1=n(ω), n2=2n(ω), and where n(ω) is given by (3). We then observe that U(t) has the form
In the frequency band where n(ω)/ω≈T0, we approximate, assuming that τ is large,
f(ωτns^)≈bf+mfωτns^,g(ωτns^)≈bg+mgωτns^.
where the slopes mf and mg, and intercepts bf and bg are found from linear least squares fits of f(x) and g(x), x∈[T0/τ,2T0/τ]. Then, all integrals in (40) can be computed explicitly.
where in the final expression we replaced n(ω) with T0ω/(2π), which is valid exactly in the high frequency regime.
References
[1]P. Agarwal, M. E. Orazem, and L. H. Garcia-Rubio. Measurement models for electrochemical impedance spectroscopy: I. demonstration of applicability. Journal of the Electrochemical Society, 139(7):1917, 1992.DOI
[2]P. Agarwal, M. E. Orazem, and L. H. Garcia-Rubio. Application of measurement models to impedance spectroscopy: III. Evaluation of consistency with the Kramers-Kronig relations. Journal of the Electrochemical Society, 142(12):4159, 1995.DOI
[3]A. Bard and L. Faulkner. Electrochemical Methods; Fundamentals and Applications. Wiley Interscience, Hoboken, NJ, 2nd edition, 2000.
[4]E. Barsoukov and J. R. Macdonald, editors. Impedance spectroscopy: theory, experiment, and applications. John Wiley & Sons Inc., 2nd edition, 2005.
[5]B. A. Boukamp. A linear Kronig-Kramers transform test for immittance data validation. Journal of the electrochemical society, 142(6):1885, 1995.DOI
[6]B. A. Boukamp. Distribution (function) of relaxation times, successor to complex nonlinear least squares analysis of electrochemical impedance spectroscopy? Journal of Physics: Energy, 2(4):042001, 2020.DOI
[7]H. Brandstätter, I. Hanzu, and M. Wilkening. Myth and reality about the origin of inductive loops in impedance spectra of lithium-ion electrodes—a critical experimental approach. Electrochimica acta, 207:218–223, 2016.DOI
[8]H.-M. Cho, Y. J. Park, and H.-C. Shin. Semiempirical analysis of time-dependent elementary polarizations in electrochemical cells. Journal of The Electrochemical Society, 157(1):A8–A18, 2009.DOI