Progress in Natural Science: Materials International xxx (xxxx) xxx

H O S T E D BY Contents lists available at ScienceDirect

Progress in Natural Science: Materials International

journal homepage: www.elsevier.com/locate/pnsmi

Data driven discovery of an analytic formula for the life prediction of

Lithium-ion batteries
Jie Xiong a, 1, Tong-Xing Lei a, 1, Da-Meng Fu a, Jun-Wei Wu a, *, Tong-Yi Zhang b, **
a
School of Materials Science and Engineering, Harbin Institute of Technology, Shenzhen, 518000, China
b
Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 511400, China

A R T I C L E I N F O A B S T R A C T

Keywords: Predicting the cycle life of Lithium-Ion Batteries (LIBs) remains a great challenge due to their complicated
Symbolic regression degradation mechanisms. The present work employs an interpretative machine learning of symbolic regression
Machine learning (SR) to discover an analytic formula for LIB life prediction with newly defined features. The novel features are
Cycle life prediction
based on the discharging energies under the constant-current (CC) and constant-voltage (CV) modes at every five
Lithium-ion batteries
cycles alternately. The cycle life is affected by the CC-discharging energy at the 15th cycle (E15
CCD ) and the
difference between the CC-discharging energies at the 45th cycle and 95th cycle (Δ45
95 ). The cycle life highly
correlates with a simple indicator ðE15
CCD
3Þ=Δ45
95 with a Pearson correlation coefficient of 0.957. The
machine learning tools provide a rapid and accurate prediction of cycle life at the early stage.

1. Introduction model to predict the long-term calendar life of LIBs as a function of

storage temperature, and found that the lithium oxidation on the nega-
Electric vehicles (EVs) have being developed extremely quickly to tive electrode is the major aging mechanism of LIBs at high temperatures.
achieve carbon neutrality and tackle climate issues of global warming Zhang et al. [10] utilized a mechanism-based model to analyze the ca-
[1]. Lithium-ion batteries (LIBs) are the power sources of EVs due to their pacity fade in LIBs used at low rates and the major capacity fade mech-
high power density, great energy density, reliable safety, and long anism was the SEI formation and the loss of active cathode material. Yang
charging-discharging cycle life compared to other commercial batteries et al. [11] presented a LIB aging model that the SEI growth was pre-
[2,3]. In general, inevitable performance deterioration of LIBs occurs and dominated at the early stage and then the lithium plating mechanism was
thus the current, voltage, temperature, charging-discharging cycle predominated later. The semi-empirical and mechanism-based models
number, state of charge (SOC), and state of health (SOH) of LIBs should have been very successful in the life prediction of LIBs, if the LIBs
be monitored to ensure reliable and safe operation [4–6]. How to use degradation follows these mechanisms.
short charging-discharging cycle information to predict the cycle life of On the other hand, data-driven models have also been developed to
LIBs still remains a considerable challenge due to the complicated investigate and analyze the aging behaviors of LIBs with the advance-
degradation mechanisms of LIBs, which have not been fully understood ment of Machine Learning (ML) techniques [12–15]. Various ML algo-
[7]. rithms including the linear model [12], random forest (RF) [16],
Great endeavors are doing and many semi-empirical and mechanism- convolutional neural network [13,16], gradient boost tree [17], Gaussian
based models have been developed to predict the cycle life of LIBs. For process regression [18], and support vector machine (SVM) [19]
instance, Bloom et al. [8] proposed a semi-empirical model, based on the demonstrated excellent predictive capability in estimating the cycle life
mechanism of solid electrolyte interphase (SEI) layer growth, to predict and Remaining Useful Life (RUL) of LIBs. Features are crucial in ML and
the accelerated calendar life and cycle life of LIBs, and the model in- how to discover features from the directly recorded information and
dicates that the impedance increase and power fade follow the derivatives, such as current, voltage, capacity, internal resistance, etc., is
square-root-of-time (i.e., time1/2) kinetics. Broussely et al. [9] built a the key point to ML. Recently, a few state-of-art ML models have

* Corresponding author.
** Corresponding author.
E-mail addresses: junwei.wu@hit.edu.cn (J.-W. Wu), mezhangt@ust.hk (T.-Y. Zhang).
1
These authors (J. Xiong and T.X. Lei) contributed equally to this work and should be considered as co-first authors.

https://doi.org/10.1016/j.pnsc.2022.12.002
Received 4 November 2022; Received in revised form 15 November 2022; Accepted 8 December 2022
Available online xxxx
1002-0071/© 2022 Chinese Materials Research Society. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article as: J. Xiong et al., Data driven discovery of an analytic formula for the life prediction of Lithium-ion batteries, Progress in
Natural Science: Materials International, https://doi.org/10.1016/j.pnsc.2022.12.002
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

predicted the RUL of LIBs using early-stage data. For example, Severson respectively. In total, the CC/CV discharging energies at the 5th-100th
et al. [12] employed the information from the first 100 cycles to predict cycle (with an interval of 5 cycles) were calculated, and thus 40 dis-
the cycle life of LiFePO4/graphite LIBs under fast charging. They found charging energies have been generated as features. The features pool is
that the statistic-based “variance” is such a crucial feature that has a provided in Supplementary Material 2.
strongly negative linear correlation with the life of LIBs and the linear
model achieved a significantly low testing error of 9.1% [12]. 2.3. Model evaluation
Obviously, there are still huge rooms to further develop ML models, in
order to predict the life of LIBs, by identifying more crucial features and The performances of regression models are evaluated by the Pearson
using interpretative algorithms, which motivates us to conduct the cur- correlation coefficient (ρ), root mean squared error (RMSE), and mean
rent work. Symbolic regression (SR) is a famous ML method, which aims absolute percentage error (MAPE)
to acquire explicitly and analytically mathematical formulas from data
sP
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
[20,21], while most ML algorithms behave like black box. When inte- n
ðby i
yÞ2
grating domain knowledge, SR is able to produce physically meaningful ρ ¼ Pi¼1 2 (3)
i¼1 ðyi
yÞ
n
formulas. The SR prediction accuracy could be very high, comparable
with other conventional ML algorithms [22,23]. Recently, SR approaches sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
have been widely employed in materials science. For instance, explicit 1X n
RMSE ¼ ðby
yi Þ2 (4)
formulas have been obtained via SR algorithms for predicting mechanical n i¼1 i
properties of low-alloy steels [24,25], glass-forming ability of metallic
glasses [26], bulk properties of ABO3 perovskites [27], and oxygen
1X n
jby i
yi j
evolution reaction activities of oxide perovskites [28], while the appli- MAPE ¼ (5)
n i¼1 yi
cation of SR in LIBs is rarely reported. Therefore, the present work uses
the data of LiFePO4/graphite LIBs [12], mines physic-based features, and
where yi , b
y i , and y are the actual value, predicted value, and the mean of
adopts interpretative SR to develop an analytic formula for the life pre-
actual values, respectively. Larger ρ values, and smaller RMSE and PAE
diction. With the newly mined physic-based features, the analytic for-
indicate more accurate regression models. A ρ value of 1, corresponding
mula given by SR exhibits excellent predictive performance.
to zero-RMSE, and zero-PAE implies a perfect fitting.
2. Methodology
2.4. Cross-validation
2.1. Data acquisition
Leave-One-Out Cross-Validation (LOOCV) is utilized here for feature
The biggest publicly available dataset for LiFePO4/graphite LIBs selection and model construction. The whole data consisting of n samples
generated by Massachusetts Institute of Technology (MIT) [12] has been is split into n folds, n-1 folds form the training set and one fold (one
utilized in the present work. The MIT dataset contains sample) forms the testing set. An ML model is trained on the training set
charging-discharging cycling data of 124 commercial A123 and tested on the testing set. This process is repeated n times, and the
APR18650M1A cells, of which the nominal capacity and nominal voltage LOOCV testing result is used as the model prediction [35]. The validation
are 1.1 Ah and 3.3V, respectively. All cells were charged by 72 different performance can be regarded as an estimation of the generalization
constant current-constant voltage (CC-CV) fast-charging policies, and ability.
then subsequently discharged with the same CC-CV discharge policy. A
cell fails if its capacity is decayed to 80% of the nominal capacity, which 3. Results and discussion
is called the failure capacity. Thus, the cycle life of a cell is defined as the
number of charging-discharging cycles until reaching the failure capac- 3.1. Feature selection
ity. Due to the variation in charging policy and the uncertainty of cells,
the cycle lives of 124 cells range from ~150 cycles to ~2300 cycles, as Four ML algorithms employed here include the random forest (RF),
shown in Supplementary Material 1 Fig. S1. extreme gradient boosting (XGB) decision tree, support vector machine
with the linear kernel (SVM-linear), and the radial basis function kernel
2.2. Feature construction (SVM-RBF). The sequential backward selection (SBS) approach [36] is
wrapped with each of the four ML algorithms to find optimized features
The current, voltage, internal resistance, and temperature of each cell of discharging energies (DEs). Since hyperparameters significantly affect
are measured during cycling in environmental chambers, and these the performance of ML models, key hyperparameters of the SVM-based
measured parameters are often utilized as features to develop ML models model (i.e., the penalty parameter C) and decision-tree-based model
for cycle life prediction and the ML models usually have poor predictive (i.e., the maximum depth of decision tree D) are considered during
powers [14]. Since the discharge voltage versus capacity curves are feature selection [35,37].
relatively sensitive to the cycle life [29–34], the discharging energy of Fig. 1(a-d) show the SBS results of SVM-linear, SVM-RBF, RF, and
each cell are calculated by Eqs. (1) and (2). XGB models, respectively. As can be seen, the SVM-RBF model performs

Z  the best, followed by the XGB model and RF model, while the SVM-linear
model performs the worst. The SBS wrapped with SVM-RBF yields the
EN
CCD ¼ VdQ (1)
N
CCD lowest LOOCV RMSE value of 121.7 cycles when C ¼ 4  104. The
corresponding selected subset contains 4 features of E10
CCD , E20
CCD ,

Z 
Z 
E65
CCD , and E90
CCD . However, the LOOCV RMSE remains almost un-
EN
CVD ¼ VdQ ¼ V dQ (2)
N
CVD N
CVD
changed (121.8 cycles) when removing E65
CCD . Therefore, the subset
consisting of E10
CCD , E20
CCD , and E90
CCD is finally selected, and fed
where V and Q represent, respectively, the voltage and capacity of the into the SVR-RBF model for further analysis.
constant current discharging (CCD) and constant voltage discharging As can be observed, only CCD energy features affect the cycle life of
(CVD) at the N-th cycle, see Supplementary Material 1 Fig. S2. For testing LIBs. The result is expected because Fig. S2 shows that the area
instance, E5
CCD and E5
CVD denote the 5th cycle discharging energies at under constant current is much larger than that under constant voltage,
constant current discharging and constant voltage discharging, meaning that the 99% discharging is finished under the constant current

2
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

Fig. 1. The LOOCV RMSE versus the feature number of each (a) SVM-linear, (b) SVM-RBF, (c) RF, and (d) XGB model with different hyperparameters during the SBS
procedure. (e) The tuning the hyperparameter γ in the SVM-RBF model (C ¼ 40000) by using LOOCV RMSE. (f) The ML predicted cycle life versus the observed one,
where the SVM-RBF (C ¼ 40000, γ ¼ 216) model is with the three features of E10
CCD , E20
CCD , E90
CCD .

mode and the constant voltage discharging lets the electrode to reach are conducted with the 20 CCD features to develop a more accurate ML
thermodynamic equilibrium. model.
Tuning the kernel parameter γ in the SVM-RBF model further im-
proves the predictive accuracy. Fig. 1(e) shows that when γ ¼ 216, the
LOOCV RMSE yields the minimum 118.19 cycles, and the corresponding 3.2. Analytic formula to predict charging-discharging cycle life
model is termed as SVM-RBF-DE model. Fig. 1(f) plots the observed cycle
life versus the LOOCV prediction of the SVM-RBF-DE model, whose The CCD features are combined by simple math operators and the
MAPE is relatively low as 19.56%. The result implies that the CCD fea- combined features are selected by the interpretive ML algorithm SR,
tures are able to catch the major of LIB life and thus feature combinations aiming to discovery a mathematical formula for LIB life. Generally, SR
utilizes Genetic Programming (GP) to search for such a formula from

3
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

formula candidates (individuals). With an initial population of random to

candidates, SR assesses all candidates and selects promising individuals.
Next generation population is constructed by fresh individuals generated A  ðE15
CCD
3Þ
Cycle life ¼ þ BðN ¼ 75; 80; 85; 90; 95; 100Þ (7)
by evolution operators including crossover and mutation. In this way, the Δ45
N
best individuals will survive and finally the best one will win the
in which Δ45
N ¼ E45
CCD
EN
CCD . Fig. 3(a) shows that the ρ value
championship of analytic formula to predict charging-discharging cycle
ðE15
CCD
3Þ
life. Please see the Supplementary Materials 1 for more details of SR. between cycle life and Δ45
N increases monotonically with N rising
With 20 CCD features and the four basic mathematic operators of from 75 to 100, while Fig. 3(b) indicates that the lowest LOOCV RMSE
addition, subtraction, multiplication, and division, the SR performs 1000 (96.9 cycles) of Eq. (7) is yielded when N ¼ 95 instead of N ¼ 100, where
iterations. Each iteration consists of 1000 individuals, and every indi- ρ ¼ 0:922. Therefore, Eq. (8) is the optimal formula balancing prediction
vidual is represented as an expression tree, with constants, features, and accuracy and model generality.
mathematic operators as nodes. The depth of trees is permitted to a
certain depth, 12 here, to limit the model complexity and thus avoid 53:17ðE15
CCD
3Þ
Cycle life ¼ þ 221:69 (8)
overfitting. Ten independent GP runs were conducted in this work, and Δ45
95
Eq. (6) with the highest ρ was obtained from one million individuals. Fig. 3(c) shows the LOOCV prediction of versus the observed life of
the LIBs. Similar to Δ45
100 , the value of Δ45
95 can accurately recognize
70ðE15
CCD
3Þ
Cycle life ¼ þ 134:76 (6) the cycle life as well, see Supplementary Material 1 Fig. S4, the cycle life
E45
CCD
E100
CCD is short than 400 cycles and larger than 700 cycles when Δ45
95 > 0:1 and
Fig. 2(a) demonstrates a robust linear correlation between cycle life Δ45
95 < 0:03, respectively.
and ðE15
CCD
3Þ
with appreciably high ρ ¼ 0.930, indicating a strong The histogram of predicted error ei ¼ b y i
yi , with predicted life b
yi
E45
CCD
E100
CCD
predictive power of Eq. (6). Fig. 2(b) shows that most cells have the and measured life yi of cell i ði ¼ 1; ⋯; 124Þ, is plotted in Fig. 3(d) and
largest CCD energy around the 15th cycle, and thus the E15
CCD appears fitted by a normal distribution, which yields the mean e ¼
0:05 cycles
as a baseline in Eq. (6). The difference in CCD energy at the 45th and and standard diviation σ e ¼ 150:64 cycles. As expected, the predicted
100th cycle (Δ45
100 ¼ E45
CCD
E100
CCD ) can reflect the degree of en- error basically follows the normal distribution and there are total four
ergy degradation at the early stage. As compared in Fig. 2(c-e), a larger data locating beyond 3σ e , as shown in Fig. 3(d). Thus, the four data are
Δ45
100 implies a more significant degradation, and leads to a shorter taken statistically as outliers [38]. Excluding the four outliers further
cycle life. Therefore, the value of Δ45
100 can be regarded as a simple enhances the prediction accuracy to LOOCV MAPE ¼ 12.1%, and LOOCV
identify of cycle life, see Fig. 2(f), the cycle life is short than 400 cycles RMSE ¼ 79.3 cycles. The four outliers are named as O1–O4, respectively,
and larger than 700 cycles when Δ45
100 > 0:1 and Δ45
100 < 0:035, and Table 1 lists their real life and predicted life.
respectively. The cycle life of 118 (95.16%) cells can be correctly clas- Generally, the discharging capacity and corresponding CCD energy
sified with the Δ45
100 criterion. tends to decline as the cycle number increases. However, the discharging
In the life prediction, it is desirable to use the information as earlier as capacity of O1-cell appearances an abnormal tendency during the 92nd
possible charging-discharging cycles. For this purpose, we extend Eq. (6) to 95th cycle, as shown in Fig. 4(a-b), and this temporary capacity

Fig. 2. (a) The linear correlation between cycle life and ðE15
CCD
3Þ=Δ45
100 . (b) The histogram of cycle with the maximum constant-current discharging (CCD)
energy. The difference in CCD energy at the 45th and 100th cycle (E45
CCD
E100
CCD ) of the cell whose life is (c) 300 cycles, (d) 480 cycles, and (e) 1155 cycles. (f)
The relationship between E45
CCD
E100
CCD and the cycle life of 124 cells.

4
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

Fig. 3. (a) The ρ value between cycle life and ðE15
CCD
3Þ=Δ45
N , and (b) the LOOCV RMSE of A ðE15
CCD
3Þ =Δ45
N þ B for predicting cycle life when N is 75,
80, 85, 90, 95, and 100. (c) The predicted cycle life of 53:17 ðE15
CCD
3Þ =Δ45
95 þ 221:69 are plotted against the corresponding measured values, where the
numbers in parentheses indicate the accuracy after excluding the four outliers. (d) The histogram of predicted errors for all 124 cells.

respectively. (Details see the Supplementary Materials 1)

Table 1
The cycle life and predictions of four outlier cells.
4. Conclusions
Outlier Cycle life (cycle) Prediction of SR (cycle)

O1 1156 1835.8 In this work, we proposed an ML framework with novel features to

O2 1801 1154.9 predict the charging-discharging cycle life of LiFePO4/graphite LIBs. The
O3 1836 1156.7
SR approach was conducted with the novel features as fundamental
O4 1935 1438.7
blocks to discover an analytic formula. The developed analytic formula is
a desirable theoretical guidance to the prediction of cycle life of LiFePO4/
recovery leads to the overestimation of cycle life for O1-cell. For this graphite LIBs with the data of first 95 cycles. The main conclusions are
outlier, using ðE90
CCD þE
2
100
CCD Þ
to replace E95
CCD enhances the predicted summarized below.
cycle life to 1005 cycles. The other three outliers might be caused by the
remarkable capacity degradation between the 45th cycle and 95th cycle, 1. The discharging energies at the constant current stage have signifi-
as shown in Fig. 4(c), for comparison, Fig. 4(d) shows most normal cant influence on the charging-discharging cycle life of LiFePO4/
degradation curves. The voltage-capacity curves of O2- and O4-cell are graphite LIBs.
provided in the Supplementary Material 1 as Fig. S5. 2. The cycle life is mainly affected by the CC-discharging energy at the
The reported performance of ML models on the same dataset is 15th cycle (E15
CCD ) and the difference between the CC-discharging
duplicated in Table 2 and the comparison indicates that the analytic energies at the 45th cycle and 95th cycle (Δ45
95 ). E15
CCD repre-
formula slightly improves the prediction accuracy. In addition, the pre- sents the maximum of the CC-discharging energy, and Δ45
95 reflects
diction accuracy can be further improved by training an XGB model the degree of energy degradation at the early stage.
(termed as the XGB-SR model) fed with E15
CCD and Δ45
95 . The LOOCV 3. The cycle life monotonically decreases with the increase in Δ45
95 ,
ρ and LOOCV RMSE of the XGB-SR model is 0.934 and 96.1 cycles, indicating that Δ45
95 is a good indicator of the degradation degree at

5
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

Fig. 4. Cycle-to-cycle evolution of discharging voltage and discharging capacity of the O1-cell (a) in the first 500 cycles, and (b) in the 90th to 95th cycle, the insight is
zoomed-in curves. The 15th, 45th, and 95th cycle evolution in the (c) O3-cell and (d) a normal cell.

Acknowledgments
Table 2
The performance of different ML models on predicting cycle life of the LIBs.
The work is supported by the Stable Supporting Fund of Shenzhen
Model Number of features RMSE (cycle) Reference (GXWD20201230155427003-20200728114835006), and the National
Elastic Net 6 127 [12] Natural Science Foundation of China (Grant Nos. 91860115).
SVM 12 115 [39]
MLP 5 119.6 [14]
Appendix A. Supplementary data
SVM-RBF 3 118.2 This work
SR 3 96.9 (79.3a) This work
a
Supplementary data to this article can be found online at https://do
without 4 outliers.
i.org/10.1016/j.pnsc.2022.12.002.

the early stage. The cycle life is shorter than 400 cycles when References
Δ45
95 > 0:1, and larger than 700 cycles when Δ45
95 < 0:03.
4. The Pearson correlation coefficient between the cycle life and the SR- [1] J. Hofmann, D. Guan, K. Chalvatzis, H. Huo, Assessment of electrical vehicles as a
proposed feature ðE15
CCD
3Þ=Δ45
95 is 0.922 and increases to 0.957 successful driver for reducing CO2 emissions in China, Appl. Energy 184 (2016)
995–1003, https://doi.org/10.1016/j.apenergy.2016.06.042.
after excluding four outliers. The analytic formula has very high [2] A. Opitz, P. Badami, L. Shen, K. Vignarooban, A.M. Kannan, Can Li-Ion batteries be
prediction accuracy with the LOOCV RMSE ¼ 79.3 cycles and MAPE the panacea for automotive applications? Renew. Sustain. Energy Rev. 68 (2017)
¼ 12.1%. 685–692, https://doi.org/10.1016/j.rser.2016.10.019.
[3] R. Schmuch, R. Wagner, G. H€ orpel, T. Placke, M. Winter, Performance and cost of
materials for lithium-based rechargeable automotive batteries, Nat. Energy 3
Declaration of competing interest (2018) 267–278, https://doi.org/10.1038/s41560-018-0107-2.
[4] L. Su, J. Zhang, C. Wang, Y. Zhang, Z. Li, Y. Song, T. Jin, Z. Ma, Identifying main
factors of capacity fading in lithium ion cells using orthogonal design of
The authors declare that they have no known competing financial experiments, Appl. Energy 163 (2016) 201–210, https://doi.org/10.1016/
interests or personal relationships that could have appeared to influence j.apenergy.2015.11.014.
the work reported in this paper. [5] L. Su, J. Zhang, J. Huang, H. Ge, Z. Li, F. Xie, B.Y. Liaw, Path dependence of lithium
ion cells aging under storage conditions, J. Power Sources 315 (2016) 35–46,
https://doi.org/10.1016/j.jpowsour.2016.03.043.

6
J. Xiong et al. Progress in Natural Science: Materials International xxx (xxxx) xxx

[6] L. Lu, X. Han, J. Li, J. Hua, M. Ouyang, A review on the key issues for lithium-ion [23] S. Sun, R.H. Ouyang, B.C. Zhang, T.Y. Zhang, Data-driven discovery of formulas by
battery management in electric vehicles, J. Power Sources 226 (2013) 272–288, symbolic regression, MRS Bull. 44 (2019) 559–564, https://doi.org/10.1557/
https://doi.org/10.1016/j.jpowsour.2012.10.060. mrs.2019.156.
[7] F. Yang, D. Wang, Y. Zhao, K.-L. Tsui, S.J. Bae, A study of the relationship between [24] Y. Chen, S. Wang, J. Xiong, G. Wu, J. Gao, Y. Wu, G. Ma, H.-H. Wu, X. Mao,
coulombic efficiency and capacity degradation of commercial lithium-ion batteries, Identifying facile material descriptors for Charpy impact toughness in low-alloy
Energy 145 (2018) 486–495, https://doi.org/10.1016/j.energy.2017.12.144. steel via machine learning, J. Mater. Sci. Technol. 132 (2022) 213–222, https://
[8] I. Bloom, B. Cole, J. Sohn, S. Jones, E. Polzin, V. Battaglia, G. Henriksen, C. Motloch, doi.org/10.1016/j.jmst.2022.05.051.
R. Richardson, T. Unkelhaeuser, D. Ingersoll, H. Case, An accelerated calendar and [25] J. Xiong, T. Zhang, S. Shi, Machine learning of mechanical properties of steels, Sci.
cycle life study of Li-ion cells, J. Power Sources 101 (2001) 238–247, https:// China Technol. Sci. 63 (2020) 1247–1255, https://doi.org/10.1007/s11431-020-
doi.org/10.1016/S0378-7753(01)00783-2. 1599-5.
[9] M. Broussely, S. Herreyre, P. Biensan, P. Kasztejna, K. Nechev, R.J. Staniewicz, [26] J. Xiong, S.-Q. Shi, T.-Y. Zhang, A machine-learning approach to predicting and
Aging mechanism in Li ion cells and calendar life predictions, J. Power Sources understanding the properties of amorphous metallic alloys, Mater. Des. 187 (2020),
97–98 (2001), https://doi.org/10.1016/S0378-7753(01)00722-4, 13–21. 108378, https://doi.org/10.1016/j.matdes.2019.108378.
[10] Q. Zhang, R.E. White, Capacity fade analysis of a lithium ion cell, J. Power Sources [27] L. Foppa, T.A.R. Purcell, S.V. Levchenko, M. Scheffler, L.M. Ghiringhelli,
179 (2008) 793–798, https://doi.org/10.1016/j.jpowsour.2008.01.028. Hierarchical symbolic regression for identifying key physical parameters correlated
[11] X.-G. Yang, Y. Leng, G. Zhang, S. Ge, C.-Y. Wang, Modeling of lithium plating with bulk properties of perovskites, Phys. Rev. Lett. 129 (2022), 055301, https://
induced aging of lithium-ion batteries: transition from linear to nonlinear aging, doi.org/10.1103/PhysRevLett.129.055301.
J. Power Sources 360 (2017) 28–40, https://doi.org/10.1016/ [28] B. Weng, Z. Song, R. Zhu, Q. Yan, Q. Sun, C.G. Grice, Y. Yan, W.-J. Yin, Simple
j.jpowsour.2017.05.110. descriptor derived from symbolic regression accelerating the discovery of new
[12] K.A. Severson, P.M. Attia, N. Jin, N. Perkins, B. Jiang, Z. Yang, M.H. Chen, perovskite catalysts, Nat. Commun. 11 (2020) 3513, https://doi.org/10.1038/
M. Aykol, P.K. Herring, D. Fraggedakis, M.Z. Bazant, S.J. Harris, W.C. Chueh, s41467-020-17263-9.
R.D. Braatz, Data-driven prediction of battery cycle life before capacity degradation, [29] D. Ansean, M. Dubarry, A. Devie, B.Y. Liaw, V.M. García, J.C. Viera, M. Gonzalez,
Nat. Energy 4 (2019) 383–391, https://doi.org/10.1038/s41560-019-0356-8. Fast charging technique for high power LiFePO4 batteries: a mechanistic analysis of
[13] L. Su, M. Wu, Z. Li, J. Zhang, Cycle life prediction of lithium-ion batteries based on aging, J. Power Sources 321 (2016) 201–209, https://doi.org/10.1016/
data-driven methods, ETransportation 10 (2021), 100137, https://doi.org/ j.jpowsour.2016.04.140.
10.1016/j.etran.2021.100137. [30] D. Ansean, M. Dubarry, A. Devie, B.Y. Liaw, V.M. García, J.C. Viera, M. Gonzalez,
[14] F. Xu, F. Yang, Z. Fei, Z. Huang, K.-L. Tsui, Life prediction of lithium-ion batteries Operando lithium plating quantification and early detection of a commercial
based on stacked denoising autoencoders, Reliab. Eng. Syst. Saf. 208 (2021), LiFePO4 cell cycled under dynamic driving schedule, J. Power Sources 356 (2017)
107396, https://doi.org/10.1016/j.ress.2020.107396. 36–46, https://doi.org/10.1016/j.jpowsour.2017.04.072.
[15] R.G. Nascimento, M. Corbetta, C.S. Kulkarni, F.A.C. Viana, Hybrid physics-informed [31] I. Bloom, A.N. Jansen, D.P. Abraham, J. Knuth, S.A. Jones, V.S. Battaglia,
neural networks for lithium-ion battery modeling and prognosis, J. Power Sources G.L. Henriksen, Differential voltage analyses of high-power, lithium-ion cells 1.
513 (2021), 230526, https://doi.org/10.1016/j.jpowsour.2021.230526. Technique and application, J. Power Sources 139 (2005) 295–303, https://doi.org/
[16] N. Yang, Z. Song, H. Hofmann, J. Sun, Robust State of Health estimation of lithium- 10.1016/j.jpowsour.2004.07.021.
ion batteries using convolutional neural network and random forest, J. Energy [32] M. Dubarry, C. Truchot, B.Y. Liaw, Synthesize battery degradation modes via a
Storage 48 (2022), 103857, https://doi.org/10.1016/j.est.2021.103857. diagnostic and prognostic model, J. Power Sources 219 (2012) 204–216, https://
[17] F. Yang, D. Wang, F. Xu, Z. Huang, K.L. Tsui, Lifespan prediction of lithium-ion doi.org/10.1016/j.jpowsour.2012.07.016.
batteries based on various extracted features and gradient boosting regression tree [33] M. Berecibar, F. Devriendt, M. Dubarry, I. Villarreal, N. Omar, W. Verbeke, J. Van
model, J. Power Sources 476 (2020), https://doi.org/10.1016/ Mierlo, Online state of health estimation on NMC cells based on predictive
j.jpowsour.2020.228654. analytics, J. Power Sources 320 (2016) 239–250, https://doi.org/10.1016/
[18] D. Yang, X. Zhang, R. Pan, Y. Wang, Z. Chen, A novel Gaussian process regression j.jpowsour.2016.04.109.
model for state-of-health estimation of lithium-ion battery using charging curve, [34] C.R. Birkl, M.R. Roberts, E. McTurk, P.G. Bruce, D.A. Howey, Degradation
J. Power Sources 384 (2018) 387–395, https://doi.org/10.1016/ diagnostics for lithium ion cells, J. Power Sources 341 (2017) 373–386, https://
j.jpowsour.2018.03.015. doi.org/10.1016/j.jpowsour.2016.12.011.
[19] A. Nuhic, T. Terzimehic, T. Soczka-Guth, M. Buchholz, K. Dietmayer, Health [35] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
diagnosis and remaining useful life prognostics of lithium-ion batteries using data- M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
driven methods, J. Power Sources 239 (2013) 680–688, https://doi.org/10.1016/ D. Cournapeau, M. Brucher, M. Perrot, E.  Duchesnay, Scikit-learn: machine learning
j.jpowsour.2012.11.146. in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
[20] C. Wang, Y. Zhang, C. Wen, M. Yang, T. Lookman, Y. Su, T.-Y. Zhang, Symbolic [36] S. Raschka, MLxtend: providing machine learning and data science utilities and
regression in materials science via dimension-synchronous-computation, J. Mater. extensions to Python's scientific computing stack, J. Open Source Softw. (2018),
Sci. Technol. 122 (2022) 77–83, https://doi.org/10.1016/j.jmst.2021.12.052. https://doi.org/10.21105/joss.00638.
[21] B. Cao, S. Yang, A. Sun, Z. Dong, T.-Y. Zhang, Domain knowledge-guided [37] T.-Y. Zhang, An Introduction to Materials Informatics, Scientific Press, Beijing,
interpretive machine learning: formula discovery for the oxidation behavior of 2022.
ferritic-martensitic steels in supercritical water, J. Mater. Informatics. 2 (2022) 4, [38] M. Panda, P.M. Khilar, Distributed self fault diagnosis algorithm for large scale
https://doi.org/10.20517/jmi.2022.04. wireless sensor networks using modified three sigma edit test, Ad Hoc Netw. 25
[22] S.-M. Udrescu, M. Tegmark, A.I. Feynman, A physics-inspired method for symbolic (2015) 170–184, https://doi.org/10.1016/j.adhoc.2014.10.006.
regression, Sci. Adv. 6 (2020), eaay2631, https://doi.org/10.1126/sciadv.aay2631. [39] Z. Fei, F. Yang, K.-L. Tsui, L. Li, Z. Zhang, Early prediction of battery lifetime via a
machine learning based framework, Energy 225 (2021), 120205, https://doi.org/
10.1016/j.energy.2021.120205.