INL/CON-07-12282

PREPRINT

Battery Technology Life

Verification Testing and
Analysis

EVS-23

Jon P. Christophersen
Gary L. Hunt
Ira Bloom
Ed Thomas
Vince Battaglia

December 2007

This is a preprint of a paper intended for publication in a journal or

proceedings. Since changes may be made before publication, this
preprint should not be cited or reproduced without permission of the
author. This document was prepared as an account of work
sponsored by an agency of the United States Government. Neither
the United States Government nor any agency thereof, or any of
their employees, makes any warranty, expressed or implied, or
assumes any legal liability or responsibility for any third party’s use,
or the results of such use, of any information, apparatus, product or
process disclosed in this report, or represents that its use by such
third party would not infringe privately owned rights. The views
expressed in this paper are not necessarily those of the United
States Government or the sponsoring agency.
 Battery Technology Life Verification Testing and Analysis

Jon P. Christophersen
1
Idaho National Laboratory
P.O. Box 1625, Idaho Falls, ID 83415
(208) 526-4280, [email protected]

Gary L. Hunt1, Ira Bloom2, Ed Thomas3, Vince Battaglia4

2
Argonne National Laboratory
9700 S. Cass Avenue, Argonne, IL 60439
3
Sandia National Laboratories
P.O. Box 5800, Albuquerque, NM 87185
4
Lawrence Berkeley National Laboratory
1 Cyclotron Road, Berkeley, CA 94720

Abstract
A critical component to the successful commercialization of batteries for automotive applications
is accurate life prediction. The Technology Life Verification Test (TLVT) Manual was
developed to project battery life with a high level of statistical confidence within only one or two
years of accelerated aging. The validation effort that is presently underway has led to several
improvements to the original methodology. For example, a newly developed reference
performance test revealed a voltage path dependence effect on resistance for lithium-ion cells.
The resistance growth seems to depend on how a target condition is reached (i.e., by a charge or a
discharge). Second, the methodology for assessing the level of measurement uncertainty was
improved using a propagation of errors in the fundamental measurements to the derived response
(e.g., resistance). This new approach provides a more realistic assessment of measurement
uncertainty. Third, the methodology for allocating batteries to the test matrix has been improved.
The new methodology was developed to assign batteries to the matrix such that the average of
each test group would be representative of the overall population. These changes to the TLVT
methodology will help to more accurately predict a battery technology’s life capability with a
high degree of confidence.

Keywords: batteries, life prediction, stress factors, Monte Carlo, uncertainty.

1. Introduction
Accurate life prediction is a critical component to the successful commercialization of high-
power battery technologies for various applications, including the automotive industry. As part
of the U.S. Department of Energy’s Advanced Technology Development Program, the
Technology Life Verification Test (TLVT) Manual [1] was developed to estimate a battery’s life
expectancy based on its anticipated usage (e.g., 15 years and 150,000 miles) at a target
confidence interval with only one or two years of accelerated aging. The manual incorporates the
FreedomCAR (Freedom Cooperative Automotive Research) goals and requirements, and can be
applied to numerous applications (minimum power assist, maximum power assist, plug-in
hybrids, ultracapacitors, etc.). This paper presents an overview of the TLVT process, followed by
some lessons learned and subsequent improvements to the methodology as a result of the
validation effort.
2. Technology Life Verification Testing
A general flow diagram of the TLVT process is shown in Figure 1. Prior to use in any TLVT-
related application, a battery technology must first be thoroughly characterized and understood.
In the absence of such knowledge, a series of short-term preparatory tests are conducted to help
identify aging mechanisms and battery stress levels. Both prior knowledge and data from such
preparatory testing are then used to develop a life model for a particular chemistry. Both a life
model and error model are necessary for a viable experiment. The life model (either empirical or
physics-based) reflects the typical degradation over time as a function of the identified life-
limiting wearout mechanisms. Common stress factors include temperature, state-of-charge,
energy throughput, and pulse power levels for both discharge and charge. The error model
accounts for both the variability due to measurement error and manufacturing differences.

A full factorial core-life test matrix (i.e., a matrix that accounts for all of the degradation
parameters identified in the life model) provides adequate coverage of these stress factors at
various acceleration rates (e.g., three temperatures and two states-of-charge) to estimate life with
a desired degree of confidence. The limiting values of these degradation mechanisms must also
be identified so that the stress level is not increased to a point where a battery’s failure rate is
different than under normal use conditions. Once the core matrix is designed, Monte Carlo
simulations are used to generate a large number of independent simulation trials for each
experimental condition using random perturbations to the model parameters, measurement error,
and manufacturing variability. The set of simulations provides a distribution of estimated average
battery life. If the Monte Carlo simulations do not yield the desired life at the desired confidence
limit (i.e., 15-year life at 90% lower confidence level), then the number of cells, and/or allocation
of cells within the core matrix has to be modified. This process is repeated until an acceptable
confidence limit is reached using Monte Carlo simulations.

The Monte Carlo simulations are then verified with actual battery testing. The performance
results are modeled and analyzed for lack-of-fit using statistical methods (i.e., bootstrap) that
create multiple realizations of the test data to assess the uncertainty in the life-on-test estimates.
If the confidence limit is acceptable and matches reasonably well with the Monte Carlo
simulation results, then the model can be used for accurate life prediction. Otherwise, the model
needs additional development, and the process then repeats. Ideally, the test matrix would
include a substantial number of batteries to minimize error, but that may not be practical or
economically feasible. Other options include testing a small number of batteries at each test
condition within the matrix, or a larger number of batteries at only a few of the test conditions.
These options will depend on the levels of manufacturing variability and measurement error as
well as the number of available batteries and test channels. For example, if the manufacturing
variability is high, then more batteries need to be placed at each test condition. After initial
characterization, both the batteries and test facilities should be checked for desired levels of
repeatability and accuracy.

A supplemental test matrix is typically executed in parallel to the core testing. This matrix is
intended to verify various assumptions about performance degradation in the model. These
assumptions may include factors such as path independence, low temperature operation, and the
effects of cold cranking. These factors would be tested to verify that they do not have an impact
on battery life. If the assumptions are proven invalid, then the model has to be re-developed, and
the core matrix simulation and testing has to be repeated.
 Prior Preparatory
Knowledge Matrix

Develop
Model

Design
Design Core
Supplemental
Matrix
Matrix

NO NO
Monte Carlo Modify Core
Simulations Matrix

Acceptable
NO
LCL?

YES

Test
Test Core
Supplemental
Matrix
Matrix

Acceptable Valid
LCL? Assumptions?

YES Done YES

Figure 1: TLVT flow diagram

3. Lessons Learned

3.1 Minimum Pulse Power Characterization

The standard method of measuring performance degradation under the FreedomCAR Program is
with a hybrid pulse power characterization (HPPC) test [2]. The profile is shown in Figure 2 and
consists of a 10-s discharge pulse (typically at a 5C1 rate), and 40-s rest period, then a 10-s charge
(more appropriately regenerative braking or “regen”) pulse (typically at a 3.75C1 rate). There is a
one-hour rest at open-circuit immediately prior to the discharge pulse to allow the battery to reach
electrochemical and thermal equilibrium. This profile is repeated at each 10% depth-of-discharge
increment (as determined from the beginning of life rated capacity), starting from a fully charged
state, and ending at a fully discharged state. In other words, a fresh battery would be subjected to
a total of nine HPPC pulses, ranging from 10 to 90% depth of discharge. From these data, the
resistance, power, and energy can be calculated using the procedures defined in the FreedomCAR
Manual [2]. For example, discharge resistance is determined using Equation (1), where Vt0 is the
open-circuit voltage immediately prior to the discharge pulse.

Vt0  Vt1
Rdis (1)
I t1  I t0

1.25
t1
1.00

0.75
Discharge
0.50
Current (relative)

0.25
t0 t2
0.00

-0.25
Regen
-0.50

-0.75
t3
-1.00
0 10 20 30 40 50 60
Time in Profile (s)

Figure 2: Hybrid pulse power characterization profile

The TLVT Manual replaced the standard HPPC test with a newly-developed minimum pulse
power characterization test (MPPC) [1]. The MPPC profile is shown in Figure 3, and consists of
an HPPC pulse profile at only two state of charge (SOC) conditions, known as SOCMAX and
SOCMIN, which are meant to cover the expected operating range of the battery while in use. (All
life testing is, therefore, also usually performed at or between these SOC conditions as well.) An
HPPC pulse profile is performed first at SOCMAX, followed immediately by a taper discharge to
the SOCMIN target condition and a one hour rest at open circuit, then another HPPC pulse profile
at SOCMIN.

1.25
SOCMAX
1

0.75
Discharge

Discharge
0.5

SOCMIN
Current (relative)

0.25

-0.25
Regen

Regen
-0.5

-0.75

Taper Discharge to SOCMIN

-1 HPPC Pulse at SOCMAX HPPC Pulse at SOCMIN
with a 1-hr Rest

-1.25

Time

Figure 3: Minimum pulse power characterization profile

Since time spent at full charge has a negative impact on lithium-ion battery performance [3], the
MPPC was designed to operate only within the specified SOC window and never let the battery
reach a fully charged state. Consequently, the batteries are charged to the specified SOCMAX
target from a lower SOC condition. However, a study with the MPPC test revealed that fast rate
charges to a specified target condition do not yield the same resistance values as when discharged
to the same SOC for lithium-ion cells [4]. Figure 4 shows the difference in resistance values
between an MPPC and HPPC test for a fresh and aged lithium-ion cell developed for the
Advanced Technology Development Program. The cell chemistry for these cells was a
LiNi0.8Co0.15Al0.05O2 positive electrode, a MAG-10 graphite negative electrode, with an
electrolyte consisting of 1.2 M LiPF6 in EC:EMC [5]. As shown in Figure 4, the resistance
values at SOCMIN (approximately 3.65 V) are similar regardless of the test profile since the
battery was always discharged to this SOC. However, at SOCMAX (approximately 4.0 V), there is
a noticeable difference in resistance values. The HPPC pulse was preceded by a discharge to
SOCMAX whereas the MPPC pulse was preceded by a charge. Table 1 shows the discharge
resistances from the MPPC and HPPC profiles at SOCMAX; the percent difference between them
are 11.5 and 5.5% for the fresh and aged cell, respectively. These data not only indicate that there
is a path-dependence effect, but also that the measured rate of resistance growth is apparently
affected by the path as well. The aged cells seem to show less path dependence than the fresh
cells. Additionally, the resistance behavior observed from the MPPC profile appears to be more
erratic than the corresponding HPPC profile. These observations were also seen in other lithium-
ion chemistries as well [4]. Obviously, changes in impedance growth have a strong impact on life
prediction, and, thus, must be measured consistently to ensure accuracy and high confidence.
These data suggest that the MPPC profile is inadequate for that purpose and should be replaced.

90

80

70
Discharge Resistance (mohms)

60

50

40

30

20

10
HPPC Discharge Pulse Resistance (Fresh Cell) MPPC Discharge Pulse Resistance (Fresh Cell)
HPPC Discharge Pulse Resistance (Aged Cell) MPPC Discharge Pulse Resistance (Aged Cell)
0
3.4 3.5 3.6 3.7 3.8 3.9 4 4.1

Open Circuit Voltage (V)

Figure 4: Discharge resistances for a fresh and aged cell (HPPC and MPPC)

Table 1. Discharge resistance values at SOCMAX from the MPPC and HPPC tests
MPPC HPPC
%-Difference
(mohms) (mohms)
Fresh Cell 32.01 35.68 11.5%
Aged Cell 62.08 65.47 5.5%

3.2 Measurement Uncertainty

Accurate life prediction requires the incorporation of both cell-to-cell manufacturing variability
and test measurement error. The original TLVT Manual estimated measurement error from the
test data, but this was found to be inadequate. A more direct approach was implemented based on
the uncertainty studies performed at the Idaho National Laboratory [6,7]. Prior to battery aging,
the test equipment is calibrated based on the manufacturer recommendations. After calibration,
accuracy measurements are taken from the test channel at various different voltage and current
levels within the full scale range using an independent digital voltmeter and calibrated shunt (to
measure current). These data are then used in conjunction with the calibration errors from the
digital voltmeter and shunt to determine the total equipment and channel error. These error
values are used to determine if the test channel is yielding data within the claimed repeatability
range (e.g., 0.02% of full scale).
These calibration data are also useful for determining the uncertainty range of derived parameters
such as resistance, power, and energy. Since the only actual measurements taken during a test are
voltage and current (temperature is treated elsewhere), the uncertainty of any derived parameter is
calculated using a propagation approach based on a sequence of Taylor Series partial derivatives
that layer down to the independent measured parameters (i.e., voltage and current). The resulting
uncertainty expression incorporates the standard deviations determined from the accuracy check,
the calibration errors, and the full-scale test channel levels. For example, Equation (2) shows the
uncertainty expression for pulse resistance, where VFS and IFS are the test channel’s full scale
voltage and current, respectively; %errVCAL and %errICAL are the calibration errors due to the
digital voltmeter and shunt from the independent measurements; and %errVSTD and %errISTD are
the standard deviations determined experimentally from the accuracy measurements.

1
2 2 2
ª § %errV · § %errI STD · º
% RS «2 ¨ STD
VFS ¸  2 ¨ I FS ¸  (%errVCAL ) 2  (%errI CAL ) 2 » (2)
«¬ © V (ta )  V (t b ) ¹ © I (ta )  I (t b ) ¹ »¼

Figure 5 shows the discharge resistance vs. open-circuit voltage curve from an HPPC test on a
fresh and aged cell (the same data as in Figure 4) with the associated upper and lower
uncertainties (i.e., dashed lines) as determined from Equation (2). The error bars are hard to see
because the measurement uncertainty is very low. Table 2 shows the uncertainty of the discharge
pulse at each 10% depth-of-discharge increment. The fresh cell shows an average error of
approximately 0.36%, and the aged cell has an average error of only 0.13%. The aged cell
uncertainty is smaller because the voltage difference (i.e., “V(ta) – V(tb)” in the denominator of
Equation (2)) is larger. Since the measurement uncertainty is so low, the data can be used for
accurate life prediction.

60 80

55 75
Fresh Cell Discharge Resistance (mohms)

Aged Cell Discharge Resistance (mohms)

50 70

45 65

40 60

35 55

30 50

25 45

HPPC Discharge Pulse Resistance (Fresh Cell) HPPC Discharge Pulse Resistance (Aged Cell)
20 40
3.4 3.5 3.6 3.7 3.8 3.9 4 4.1

Open Circuit Voltage (V)

Figure 5: Discharge resistances for a fresh and aged cell (HPPC and MPPC)
Table 2. Measurement uncertainty for a fresh and aged cell
Fresh Cell Aged Cell
Open-Circuit Discharge %-err Open-Circuit Discharge %-err
Voltage (V) Resistance (m:) Voltage (V) Resistance (m:)
3.99 35.68 0.34% 4.00 65.47 0.12%
3.90 35.16 0.35% 3.91 60.64 0.13%
3.82 33.23 0.37% 3.82 58.51 0.13%
3.73 30.52 0.40% 3.73 57.04 0.14%
3.65 30.30 0.40% 3.66 57.13 0.14%
3.60 31.37 0.39% 3.61 61.47 0.13%
3.55 33.17 0.37% 3.56 76.25 0.11%
3.47 46.06 0.27%
Average: 0.36% Average: 0.13%
Standard Deviation: 0.04% Standard Deviation: 0.01%

3.3 Rank Ordering of Cells

Another important aspect to accurate life modeling is the appropriate allocation of cells into a test
matrix based on manufacturing variability. Ideally, manufacturing variability will be sufficiently
low such that the cells can be distributed into a matrix at random, but this is not yet practical or
economically feasible. Consequently, a methodology was developed to “randomly” assign cells
to the matrix such that the average of each test group is representative of the overall population
with respect to resistance and capacity. The deviation in resistance and capacity from the average
for each cell is determined using Equations (3) and (4), respectively, where “Q” is cell capacity,
and “R” is cell resistance. The cells are then randomly assigned to each group within the matrix
such that the overall group average deviation within a test condition is less than or equal to some
predetermined limit using the process shown in Figure 6. Once a deviation limit is chosen (i.e.,
0.5), a random sample of cells is chosen to match a particular matrix condition such that there are
no duplicates (i.e., the same cell number is chosen twice) or repeats (i.e., a cell that has already
been placed in another matrix condition). The average deviation in resistance and capacity are
then determined (i.e., the average of “Rdev” and “Qdev” for each cell in the matrix condition). If
the averages are below the deviation limit, then the cells are kept in the matrix condition.
Otherwise, they are placed back into the pool of available cells again, and the process repeats
until all the matrix conditions have been filled. Figure 7 shows an example rank ordering of cells
based on a simple random assignment (squares) and a random assignment within a deviation limit
of 0.5 (diamonds). As shown, the group averages for the rank ordering of cells are much tighter
with a deviation limit.

Rcell  Ravg
Rdev (3)
Rstd

Qcell  Qavg
Qdev (4)
Qstd
Figure 6: Rank ordering of cells flow diagram

0.6
Rank Ordered Cell Grouping
Purely Random Cell Grouping
0.4

Rank Ordering
C/1 Capacity at 30C (STD from AVG)

0.2 Deviation Limit of 0.5

-0.2

-0.4

-0.6

-0.8

-1

-1.2
-1 -0.5 0 0.5 1 1.5 2

Discharge Resistance at SOCmax (STD from AVG)

Figure 7: Sample rank ordering of cells with a deviation limit of 0.5

4. Future Work

The TLVT Manual validation effort is continuing, and there are other aspects of the methodology
that need to be investigated. Of particular importance is a generalized empirical degradation
model that can be applied to a set of data in the absence of a manufacturer’s mechanistic model.
The generalized model has the form shown in Equation (5), where “P” is the average response in
the battery population (e.g., resistance), “Xi” is the level of the ith stress factor (e.g., temperature,
SOC, etc.), and the “Ei” and “U” terms are model parameters. The error model is shown in
Equation (6), where Var(Y(X1,X2,…,Xq;t)) is the variance of the response (“Y”) at time “t” for
batteries stressed by the experimental condition specified by {X1, X2,…,Xq}. The variance of a
random cell-to-cell proportional effect (induced by manufacturing variability) is given by VG², and
VH² is the measurement error variance (which can now be determined independently, as discussed
above). This model is still under development, and will be part of a software tool package that
will accompany the next version of the TLVT manual.

P X 1 , X 2 , , X q ; t 1  exp^E 0  E1 ˜ X 1  E 2 ˜ X 2    E q ˜ X q `˜ t U (5)

Var Y X 1 , X 2 ,, X q ; t V G2 ˜ P X 1 , X 2 ,, X q ; t  1 2  2 ˜ V H2 (6)

Also of interest is the path dependence effect due to aging. Path dependence is applicable to a
mix of calendar (i.e., shelf-life) and cycle life aging, thermal cycling, or SOC swings. These
effects must be determined experimentally, and the life model should account for them as well.
Lastly, the validation effort to date has centered on only one lithium-ion chemistry. Other
chemistries will also be addressed to further validate the methodology and the general form of the
life model.

5. Summary and Conclusions

The TLVT methodology is a useful tool for predicting battery life with a high degree of statistical
confidence. The test matrices (core and supplemental) are statistically designed with a suggested
allocation of cells based on the stress factors and levels, number of batteries available for test, and
number of test conditions. Using an empirical or physics-based model, battery life is estimated
using Monte Carlo simulations of the various test conditions in the core matrix. Once the
simulation is satisfactory, batteries are placed on test to verify accuracy of the life model and the
assumed levels of random variability from manufacturing and testing. Also, the statistical
methods such as the bootstrap are used to assess the uncertainty in the life-on-test estimates
derived from the experimental data.

A validation effort for this manual has yielded a number of lessons learned and methodology
improvements. Testing with the newly developed minimum pulse power characterization test
showed a path dependence effect on voltage when a target condition was approached by a charge
or a discharge. The resistance growth behavior is therefore also impacted by the path dependence
as well. Consequently, the MPPC test is considered inadequate for TLVT applications and needs
to be replaced. Another critical component of accurate life prediction is a low measurement error
during aging. The Taylor Series derivation of parameters such as resistance based on calibrations
and independent checks appears to give the most accurate uncertainty range, and thus provides
the most accurate life estimation. Finally, the appropriate allocation of cells into a core or
supplemental matrix is highly dependent on manufacturing variability. If the manufacturing
variability is too high, a new approach to randomly assign cells to a test matrix is developed such
that the group average is representative of the entire set of batteries in the matrix based on
deviations from the average in both capacity and resistance.

Additional work is underway on the development of a general life model that accounts for the
identified stress factors and is capable of separating the effects of measurement error (determined
independently) and manufacturing variability from the test data. Path dependence studies that
include aging (calendar or cycle life), temperature, or state of charge are also being considered.
Lastly, a TLVT life prediction using various lithium-ion chemistries is also vital for validating the
general model.

6. References
[1] Advanced Technology Development Program for Lithium-Ion Batteries: Battery Technology Life
Verification Test Manual, INEEL/EXT-04-01986 (February 2005).
[2] FreedomCAR Battery Test Manual for Power-Assist Hybrid Electric Vehicles, DOE/ID-11069, Oct.
2003.
[3] J. P. Christophersen, C. D. Ho, C. G. Motloch, D. Howell, and H. Hess, J. Electrochem Soc., 153 (2006)
A1406.
[4] J. P. Christophersen, G. L. Hunt. C. D. Ho, and D. Howell, J. Power Sources, accepted.
[5] Advanced Technology Development Program for Lithium-Ion Batteries: Gen 2 Performance Evaluation
Final Report, INL/EXT-05-00913, Jul. 2006.
[6] Uncertainty Study of INEEL EST Laboratory Battery Testing Systems Volume 1: Background and
Derivation of Uncertainty Relationships, INEEL/EXT-01-00505 (December 2001).
[7] Uncertainty Study of INEEL EST Laboratory Battery Testing Systems Volume 2: Application of Results
to INEEL Testers, INEEL/EXT-01-00505 (March 2003).

7. Acknowledgements
This work was prepared as an account of work sponsored by an agency of the United
States Government under US DOE Contract DE-AC07-05ID14517. Funding for this
work was provided by the U.S. DOE Office of FreedomCAR and Vehicle Technologies.

8. Authors
Jon P. Christophersen
MS, 2005, Electrical Engineering, University of Idaho
Research Engineer, Energy Storage and Transportation Systems
Idaho National Laboratory
P.O. Box 1625, Idaho Falls, ID 83415, USA

Gary L. Hunt
Ph.D., 1972, Electrical Engineering, University of Texas
Consulting Engineer to Energy Storage & Transportation Systems
Idaho National Laboratory
P. O. Box 1625, Idaho Falls, ID 83415, USA
Ira Bloom
Ph.D., 1983, Inorganic Chemistry, University of Chicago
Chemist/Manager, Electrochemical Analysis and Diagnostics Laboratory
Argonne National Laboratory
9700 S. Cass Avenue, Argonne, IL 60439, USA

Ed Thomas
Ph.D., 1988, Statistics, University of New Mexico
Statistician, Independent Surveillance Assessment and Statistics
Sandia National Laboratories
P.O. Box 5800, Albuquerque, NM 87185

Vince Battaglia
Ph.D., 1993, Chemical Engineering, The University of California, Berkeley
Principle Investigator and Program Manager
Lawrence Berkeley National Laboratory
1 Cyclotron Road, Berkeley, CA 94720