Fast and direct diagnosis of states of health for the spent batteries

Fast and direct diagnosis of states of health for the spent batteries | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Fast and direct diagnosis of states of health for the spent batteries Minjeong Gong, Yoonjung Choi, Han Mo Yang, Sang Bok Ma, Dong-Hwa Seo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7524783/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Fast diagnosis of state of health (SoH) is essential for a second life of the spent batteries due to drastically increasing demands for lithium-ion batteries. However, additional voltage or state of charge (SoC) settings were needed to predict SoH. Here, we suggest a SoH prediction process without any additional settings within 110 seconds (maximum 290 seconds) using machine learning, which consists of 2 steps: protocol classification and SoH regression. Data were generated from 18650-sized cells. There are 3 protocols depending on SoC due to cut-off voltage, and open-circuit voltage and 1.0 C-rate discharge voltage for 10 seconds were collected at every 5% SoC in 188 SoH. The protocol classification model showed 0.9985 of accuracy. Then, protocol data, with current pulses and rest times, corresponding to SoC were collected at every 5% SoC in 1043 SoH, and 12 features were selected from 60 features. The SoH regression model showed 0.850% mean absolute error (MAE), and performance in cases of untrained SoC and misclassification was also superior because the model was trained on all SoC data. In addition, the suggested SoH prediction process can apply to untrained cells, such as different form factors. The performances of 21700-sized cells were improved to 2.600% MAE through feature engineering to the resistance-related features using the ratio between cell capacity and external volume. This paper highlights a fast and direct SoH prediction process without any settings, and feature engineering of untrained cell data to save time and energy for diagnosis of spent batteries. Artificial Intelligence and Machine Learning SoH Machine learning spent battery lihtium-ion battery Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction The demand for lithium-ion batteries is increasing continuously as electric vehicles (EVs) are commercialized to reduce greenhouse gas (GHG) emissions from internal combustion engines (ICEs). Thus, the number of spent batteries from EVs will also increase drastically in the near future. However, GHG emissions occur during the production of new batteries from material synthesis to battery manufacturing, although EVs can reduce GHG emissions by substituting ICEs during driving. 1 Thus, if the spent batteries can be used in energy storage systems (ESSs) or low-power EVs as a second life of batteries, it can effectively reduce GHG emissions from battery production. The states of the spent battery should be checked for evaluation of suitability for a second life. The important one among them is the state of health (SoH) that represents the remaining capacity relative to capacity at the beginning of life (BoL). To measure the SoH, it takes at least 6 hours per cell with a 0.33 C-rate for full charge and discharge. Thus, it takes considerable time to measure SoH for all the spent batteries. To quickly diagnose SoH of the spent batteries, researchers tried to predict SoH by using machine learning (ML), and there are two approaches to collect data for feature generation from galvanostatic measurements, such as voltage changes during constant current (CC) or rest time after current. The first approach was voltage data collection during CC charge, and it needed all or partial voltage profiles during charging to generate features for SoH prediction. 2 – 4 To do this, it needs additional time and energy to set a fully discharged state or a specific voltage. Another approach was collecting voltage data at a specific state of charge (SoC). The voltage changes during rest after fully charging (100% SoC) or repeating current pulses and rest times at a specific SoC were used as features. 5 , 6 In the latter case, it is important to set an accurate specific SoC due to various resistances depending on SoC. 7 Both approaches require additional pre-settings to get data for feature generation. On the other hand, Park et al. reported SoH prediction without any pre-setting by regenerating the empty region in the entire charge and discharge. 4 However, the performance was largely dependent on the amount of collected voltage profile, that if the time for voltage measurement was short, the performance of SoH prediction drastically decreased. To solve these problems, we suggest a SoH prediction process without any settings such as voltage or SoC. At this time, different protocols are applied to the spent batteries depending on the SoC near the upper or lower voltage cut-off, and it is determined based on voltages during the 1.0 C-rate discharge process for 10 seconds. Based on data collected from protocols containing various current pulses and rest times, numerous features are generated, and a trained ML model predicts the SoH of the spent cell. The overall process is conducted within 110 seconds (maximum 290 seconds), 10 seconds for protocol classification and 100 seconds (maximum 280 seconds) for protocol data, and it showed high performance for direct SoH prediction without any voltages or SoC settings. Results SoH prediction process The overall process of SoH prediction for the spent batteries with unknown SoC was shown in Fig. 1 ., and there were two parts, protocol classification and SoH regression. Because voltage reached the cut-off voltage at high and low SoC when high current was applied, a different protocol was applied to get voltage data for SoH prediction. There were three types of protocol depending on the SoC (low, middle, and high), and voltage data measured from 1.0 C-rate discharge for 10 seconds were used for classification of protocol. Then, the SoH regression model predicted SoH based on the collected data applied a classified protocol containing several current pulses and rest times. Data was generated by using 18650 cylindrical cells (INR18650-29E) with 2.85 Ah of standard capacity. The 3 protocols were applied as shown in Table 1 – 3 and Supplementary Fig. 3 . The number and boundary of protocols were determined based on the voltage profile, and the details will be discussed later. To train the protocol classification model, open-circuit voltage (OCV) and voltage data of 1.0 C-rate discharge for 10 seconds were collected with 5% SoC reduction from 100 to 0% SoC, and measured at 188 various SoH (Fig. 2 . (a-b) ). For each collected voltage data, 8 features such as OCV, V 0s V 10s , R 0 − 0.1s , R 0 − 10s , R 0.1−10s , skew, and kurt were generated (Fig. 2 . (c) ). On the other hand, to train the SoH prediction regression model, firstly, the discharge capacities of degraded cells with various cycle test conditions were measured with 0.5 C-rate charge and discharge to calculate SoH ( Supplementary Table 1 ). Then, voltage data were collected at 5% SoC interval depending on the corresponding protocol that was determined based on the SoC (Fig. 2 . (d) ). As shown in Fig. 2 . (e) , protocol data from 0 to 100% SoC were collected in 1043 various SoH. Due to getting data from various C-rate and current direction at each SoC and SoH, it was necessary to express their changes as features. Thus, features were generated to express voltage changes, kurtosis, and skewness for each C-rate (1.65 and 2.0C), pulse/rest (P/R), current direction (charge/discharge as ch/dch). Additionally, relations between C-rates or current directions, such as voltage difference after current pulse (hyst) and resistance ratio (R_ratio), were also used as features, and finally, a total 60 features were generated (Fig. 1 . (f) ). Protocol classification As shown in Fig. 3 . (a) , the lower current (0.1 C) was applied in charge for higher 80% SoC and discharge for lower 10% SoC because these SoC regions were too close to cut-off voltage ( Supplementary Fig. 3 ). Thus, it should be classified which protocol is applied to the spent cell, and the classification was conducted by using OCV, R 0 − 0.1s , R 0.1−10s , and skew as features that were selected after removing high correlation between features. Among various ML models, LGBMClassifier showed the highest accuracy as 0.9997 for train and 0.9985 for test data (Fig. 3 . (b) and Supplementary Table 2 ). At this time, there were misclassified cases at the near boundary between protocols with very low possibility, and the effects of protocol misclassification on SoH prediction will be discussed later. The SHAP (Shapley Additive exPlanations) analysis was employed to understand how to classify protocols from 4 features, and can analyze the impact of each feature to predict the target value. 8 The SHAP value means that the larger the SHAP value, the predicted value increases, and vice versa. In the case of a classification model, the SHAP value is calculated for each probability of classifying a certain class. Thus, the high SHAP value means the probability of classifying a certain class increases with a large impact. As shown in Fig. 3 . (c-f) , OCV showed the highest impact for the protocol classification. OCV is directly related to the SoC, thus, low, middle, and high OCV contribute to increasing the probability of prediction with protocol 1, 2, and 3. 9 R 0 − 0.1s represented the voltage difference between OCV and voltage at 0.1 seconds after the current pulse and was related to the ohmic resistance, containing electrical conductivity. 10 , 11 Because the electrical conductivity of NCA and graphite at the low SoC was lower than the middle and high SoC, the high R 0 − 0.1s contributed positively to protocol 1. 12 , 13 And, R 0.1−10s was directly related to the voltage difference during the current pulse for 10 seconds. Due to the large voltage drop at the low SoC, the high R 0.1−10s positively contributed to the prediction. Oppositely, low R 0.1−10s contribute positively to protocol 2 because the slope of voltage profile was less steep than the low SoC. 9 Similarly to R 0.1−10s , the high value of skew feature contributed positively to protocol 3, and in contrast, low skew contributed positively to protocol 1. Because voltage plateaus at high SoC and large voltage drops at low SoC, skewness at high and low SoC showed high and low values, and it affected the prediction. SoH regression To predict SoH, the ML model trained on protocol data measured at all SoCs in 1043 SoH and generated 60 features as shown in Fig. 1 . (e-f) . Among various ML models, RandomForest regressor showed the best performance to predict SoH ( Supplementary Table 3 ). Model optimizations, such as feature selection and hyperparameter tuning, were conducted by using the RandomForest regressor, and finally, 12 features were selected ( Supplementary Table 4 ). In Fig. 4 . (a) , the optimized SoH prediction model showed high performance regardless of protocols with 0.998, 0.984% of R 2 score, 0.232, 0.850% of MAE, 0.414, 1.267% of RMSE, and 0.287, 1.055% of MAPE for the train and test data, respectively. The SoH prediction model also showed high performance, with 2.235% of MAE, 2.741% of RMSE, and 2.799% of MAPE, from the data measured at the untrained SoC because SoH prediction model was trained on data measured at all SoC regions (Fig. 4 . (b) ). In addition, the performances where data was measured from the misclassified protocol were also confirmed (Fig. 4 . (c) ). The generated features were not similar to the train data due to different protocols, thus, RMSE, MAE, and MAPE slightly increased to 3.871% of MAE, 4.586% of RMSE, and 4.624% of MAPE. However, the accuracy for classification of protocol was very close to 1, thus, these cases will be rare, as shown in Fig. 3 . (b) . The resistance of cells varies depending on their SoC due to different ionic/electronic conductivity or reaction mechanisms. 7 Thus, in most studies, ML models were trained on data measured at a specific SoC or certain voltage regions 2 , 3 , 5 , 6 . On the other hand, our SoH prediction model was trained on features containing resistance measured in all SoCs. As a result, the SoH prediction model predicted with high accuracy, even though it was untrained on data such as SoC or protocol misclassification. The SHAP analysis was conducted to understand how SoH prediction model was trained, and 3 features with the highest mean absolute SHAP value were visualized in Fig. 4 . (d-f) . These features were related to resistance and highly correlated with SoH ( Supplementary table 4 ). The SHAP values of these features decreased as feature values increased due to increasing resistances with cell degradation. These 3 features also varied depending on the SoC, similar to the previous study ( Supplementary Fig. 4 ). 7 The 1.65C_dch_P_R 0 − 0.1s , which was the highest correlation with SoH, was the most evenly distributed in all SoC ranges. However, the resistance at 100% and below 10% SoCs showed high resistance because resistances were calculated from the voltage difference, and large voltage drops were observed at those regions. Therefore, it should be compensated by using other features to increase accuracy at 100% and below 10% SoCs. First, 2.0C_ch_R_R 0 − 0.1s at 100% SoC showed much lower values compared to other SoCs because a constant voltage of 4.2 V was applied. Thus, the scaled 2.0C_ch_R_R 0 − 0.1s between 0 and 0.25 at protocol 3 were measured at 100% SoC, and the SHAP values showed different tendencies as shown in Fig. 4 . (e) . On the other hand, 1.65C_dch_P_R 0.1−5s were selected to increase accuracy below 10% SoCs. As shown in Supplementary Fig. 4. (c) , 1.65C_dch_P_R 0.1−5s was observed in a wide range due to abrupt voltage drop at low SoCs. Thus, it could accurately predict SoH values, although higher 1.65C_dch_P_R 0 − 0.1s than other SoCs. Expandability of SoH prediction process To confirm expandability, the suggested SoH prediction process was applied to the other cells. 2 types of 21700-sized cylindrical cells with different electrode chemistry and standard capacity were used (Table 4 ), and their voltage profile and dV/dQ were shown in Fig. 5 . (a, b) . The voltage drops at the low SoC of cell 2 started at a higher SoC due to incorporation of Si in the anode. And the peak near SoC 80% in dV/dQ of cell 3 was shifted to slightly lower SoC due to different cathode chemistry. Because the values of generated features showed a large difference due to the different form factors of cell 2 and 3 compared to cell 1, it needs feature engineering for scaling from 21700 to 18650-sized cells to predict SoH by using SoH regression model trained on data from cell 1. Largely contributed features were related to resistance, such as dV 0 − 0.1s and dV 0.1−10s for protoco1 classification and 1.65C_dch_P_R 0 − 0.1s , 2.0C_ch_R_R 0 − 0.1s , and 1.65C_dch_P_R 0.1−5s for SoH regression. Thus, we concentrated on the features related to the resistance for feature engineering, such as rescaling the resistance of different form factor cells to cell 1. There are numerous components affecting resistance, such as ionic/electronic conductivity, loading level, and electrode density. However, it was difficult to identify the exact information of the spent batteries, such as active material chemistry, loading level, and electrode density without additional analysis after cell disassembly. Thus, feature engineering without any detailed information was needed, and the ratio between cell capacity and external volume, such as 18650 and 21700 size, was used for feature engineering. As shown in Fig. 5 . (c) , the ratio of cell capacity and external volume showed a linear relationship with cathode and anode loading levels. When the loading level was higher, the resistance of the cell increased due to a longer diffusion path and higher tortuosity. 14 , 15 Thus, the features related to resistance were rescaled based on the ratio of cell capacity and external volume. To classify protocols for other cells, feature engineering was conducted on the features related to resistance using the ratio between cell capacity and external volume. Because the resistances were not directly related to the ratio between cell capacity and external volume, though it showed a linear relation with loading level, slope and intercept optimization were conducted on the features related to resistance, dV 0 − 0.1s and dV 0.1−10s . The optimized equations of each feature for rescaling from 21700 to 18650-sized cell were shown in Supplementary Fig. 5 . After feature engineering to dV 0 − 0.1s and dV 0.1−10s , it showed a similar tendency to cell 1. Thus, the accuracy also increased from 0.939 to 0.963 for cell 2, and the accuracy at low SoC was especially improved ( Supplementary Fig. 6 ). However, different active materials chemistries caused accuracy reduction compared to cell 1, although feature engineering was performed. Thus, because these different chemistries were used in the cells and OCV, which largely dependent on chemistry, was the most effective feature for protocol classification, 10% SoC in cell 2 and 75% SoC in cell 3 were still misclassified ( Supplementary Fig. 7 ). In case of cell 2, graphite with Si was used as an anode material, and it caused the voltage drop between 0% and 20% SoCs. In contrast, the NCM as a cathode material was used in cell 3 compared to NCA in cell 1. Thus, the peak position in dV/dQ of cell 3 at near 80% SoC was slightly lower than cell 1, causing misclassification at 75% SoC. Then, protocol data were collected based on the SoC for SoH prediction, and SoH predictions for cell 2 and 3 were performed by using a SoH regression model trained on data from cell 1 data. And feature engineering was also conducted to feature related to resistance, similarly to protocol classification. Among various resistance-related features, the 3 features with the highest mean absolute SHAP value, 1.65C_dch_P_R 0 − 0.1s , 2.0C_ch_R_R 0 − 0.1s , and 1.65C_dch_P_R 0.1−5s , were selected for feature engineering to evaluate the prediction performance for the other cells ( Supplementary Fig. 8–10 ). After feature engineering, the MAE of cell 2 and 3 were reduced from 4.230% and 4.768–2.602% and 2.589%, respectively (Fig. 5 . (e) and Supplementary Fig. 11 ). Similar to the protocol classification results, large errors were observed at the different reaction occurred due to chemistry differences (black-edged marker in Fig. 5 . (e) ). Discussion In this study, SoH of the spent battery was predicted within 110 seconds (maximum 290 seconds) with short current pulses and rests to classify protocols and predict SoH. As shown in Fig. 5 . (f) , there were several studies to predict SoH of the spent batteries. Most studies concentrated on the data measured at specific voltage ranges, thus, voltage setting was necessary to predict SoH. 2 , 3 , 5 , 16 Or it needed SoC setting to predict SoH because data collection and training model were conducted at a specific SoC in previous studies. 5 , 6 On the other hand, it was reported that SoH was predicted by regeneration of the empty region in the entire voltage profile, thus, no settings were needed. 4 However, the performance was critically affected by the amount of empty region. Thus, it took several hours to improve the performance. To address these problems, we suggested SoH prediction process regardless of SoC of the spent batteries without any SoC and voltage setting by training on all SoC range data to the regression model for SoH prediction. The suggested process could be used to predict SoH directly with high accuracy and very short consumption time for SoH diagnosis. In addition, other cells with different form factors can also be predicted through feature engineering. Feature engineering was conducted to only resistance-related features by using minimum cell information, such as the ratio between cell capacity and external volume that relates to the electrode loading level, and it effectively improved the performance of protocol classification and SoH regression. SoH of the spent batteries for a second life can be determined within a short time through the suggested SoH prediction process with protocols and ML models, and energy and time to diagnose SoH for the spent battery will be saved. Furthermore, in this study, feature engineering was only conducted on the feature related to resistance, assuming electrode material chemistries were unknown. If it is possible to know the exact chemistries of the electrode material, feature engineering for other features can also be conducted with a suitable method to improve performance. The SoH of other form factors and chemistry cells can be predicted through feature engineering without additional data collection and training ML models, thus, energy and time are also saved to collect additional data of various cells. 6 Conclusion In this study, 3 types of protocols containing various current pulses and rest time were applied, and ML models predicted a suitable protocol and SoH for the spent batteries with unknown SoC. The protocol classification model showed a high accuracy of 0.9985 for test data. SoH regression model predicted SoH based on the voltages measured through the corresponding protocol and showed a high R 2 score of 0.984 for the test data. The SoH prediction model also showed high performance in case of misclassification and untrained SoC by training on data in all SoCs. In addition, the performance of other cells with different form factors was improved through feature engineering to resistances-related features using the ratio between cell capacity and external volume, even though these cell data were not trained. The performance of SoH predictions will be improved through feature engineering reflecting the chemistry of electrode materials. Method Data generation Commercial cylindrical lithium-ion battery (LIB) cells, INR18650-29E (NCA/Graphite, Samsung SDI), were used for data generation. The chemistry and standard discharge capacity of the cell were summarized in Table 4 . We used a 5 V and 12 A maximum cycler (BTS-4008-5V12A-S1, Neware) for the cell cycle test, protocol classification test, discharge capacity check, and three types of protocol tests. All the LIB cells were tested in a chamber maintained in room temperature (RT). Each LIB cell was cycled in the voltage range of 2.5–4.2 V with different charge and discharge C-rate conditions that are summarized in Supplementary Table 1 : a constant current (CC) / constant voltage (CV) mode for charging step (end condition: 0.02 C-rate cut-off) and a CC mode for discharge step at RT. The sample size of each cycle test was 3 EA except for the test condition with 4.0 C-rate discharge, 6 EA. The protocol classification test, the discharge capacity check and protocol test were performed under the adjustment of cycle number to get data from evenly distributed SoH in SoH regions we interested in ( Supplementary Fig. 1 ). In protocol classification test, the voltage data during 1.0 C-rate discharge and rest time for 1 hour after discharge were collected at 5% depth of discharge (DoD) intervals from state of charge (SoC) 100–0%. Then, the discharge capacity was measured at 0.5 C-rate (CC mode) to check SoH, and the protocol tests were conducted, that divided by three different SoC ranges: 0, 5% SoC (Protocol 1, Low SoC), 10 to 75% SoC (Protocol 2, Middle SoC), and 80 to 100% SoC (Protocol 3, High SoC). In the protocol test, the cells were discharged at 1.0 C-rate, 5% DoD intervals from 100 to 0% SoC to set the specific SoC we were interested in. After setting SoC, they were conducted with protocol 3 to 1 with two different C-rates, 1.65 C-rate and 2.0 C-rate for 5 seconds and rest time for 20 seconds to collect the voltage data for SoH regression model. In protocol 1 and 3, we applied 0.1C-rate current in discharge and charge steps, respectively, because of the cut-off voltage condition. We summarized each protocol test condition according to three different SoC ranges in Table 1 – 3 ( Supplementary Fig. 3) . To confirm the performance of SoH prediction model at untrained SoCs, the protocol tests data were collected with three different initial SoC states, 98%, 97% and 96% SoC in various SoH states of 18650-29E cells. The cells were fully charged with 1.0 C-rate in CC/CV mode (end condition: 0.02 C-rate cut-off) and then discharged with 1.0 C-rate under the end conditions of 2%, 3% and 4% DoD based on the discharge capacity we checked, respectively. There were the same as the trained SoC cases in the whole protocol test conditions according to the SoC ranges and 5% DoD intervals except for the end SoC states, 3%, 2% and 1% for 98%, 97% and 96% SoC in initial states, respectively. For the misclassified protocol test, we artificially applied the upper and lower levels of the protocol condition in Table 1 – 3 after setting the particular SoC in the cells with various SoH states. We divided four cases according to SoC values, 5%, 14%,75% and 84% SoC, respectively. In case 1 (5% SoC), the cells were fully charged and then discharged with 1.0 C-rate under the end-condition of 95% DoD based on the discharge capacity we checked. After setting 5% SoC, we conducted the test with protocol 2 (middle SoC), not protocol 1, to check the performance of SoH prediction model. In case 2 (14% SoC), case 3 (75% SoC) and case 4 (84% SoC), the SoC setting method was the same as case 1, and the protocol test were conducted in each case: case 2 (14% SoC) with protocol 1 (low SoC), case 3 (75% SoC) with protocol 3 (high SoC), and case 4 (84% SoC) with protocol 2 (middle SoC), respectively. For verification of SoH prediction process suggested, two commercial 21700-sized LIB cells, INR21700-50E (NCA/Graphite + Si, Samsung SDI) and INR21700-M50LT (NMC/Graphite, LG Chem), were used to obtain cell data. The chemistry and standard discharge capacity of these cells are summarized in Table 4 . In the case of 21700-sized LIB cells, we conducted the cycle test with 0.5 C-rate charge (CC/CV mode) / 1.0 C-rate discharge (CC mode) at RT. The sample size of each cycle test was 3 EA. We collected 1.0 C-rate discharge data for protocol classification, discharge capacity, and protocol data every 100 cycles under the same condition and environments as INR18650-29E. Data preprocessing From the data collected at 0.5 C-rate for protocol classification and SoH regression, the SoC and SoH were calculated using the following equations. $$\:SoC\:\left[\%\right]=\:\frac{Remaining\:capacity}{Fully\:discharge\:capacity}\times\:100$$ $$\:SoH\:\left[\%\right]\:=\:\frac{Discharge\:capacity}{Discharge\:capacity\:of\:BoL}\times\:100$$ Then, features were calculated from voltage data for protocol classification and SoH regression after discharge capacity measurement. In case of the protocol classification model, 3 types of values were generated, such as actual voltages, resistances, and distribution of voltage during discharge. The actual voltages were open circuit voltage (OCV), V 0.1s , and V 10s (V x s is voltage after x seconds during 1.0 C-rate discharge). At this point, quasi-OCV was used as OCV, which was the end voltage after 1-hour rest time. Resistances were calculated as follows: $$\:{R}_{x-y\:s}=\:\frac{\left|{V}_{x\:s}-{V}_{y\:s}\right|}{current}$$ where V x s was the voltage at x seconds during 1.0 C-rate discharge, and V 0s was OCV. The distribution of voltage during discharge was characterized by skewness and kurtosis of voltages for a 10-second 1.0 C-rate discharge. In the case of SoH regression model, the generated values were similar to the protocol classification model, such as actual voltages, resistances, and distributions of voltage during pulse or rest. Additionally, the relations between C-rate or current direction were also expressed as values because various current pulses and rests were contained in protocols. V 0.1s and V 5s as actual voltages, R 0 − 0.1s and R 0.1−5s as resistances, and skewness and kurtosis as distributions of voltage were generated to express each current pulse or rest. And end voltage difference at each current pulse as hysteresis (hyst), resistance ratio between charge and discharge at the same C-rate as R_ratio, and resistance ratio between 1.65 and 2.0 C-rate were generated as features to express the relation between C-rate and current direction. The features for SoH regression model were numerous, thus, features were labeled as (C-rate)_(charge/discharge)_(pulse/rest)_(value) for each pulse and rest. Machine Learning To classify protocol and predict SoH by using machine learning (ML), feature selection and model optimization, such as model selection and hyperparameter tuning, should be conducted. The same procedures were conducted for protocol classification and SoH regression model. Because there were numerous generated features, 8 features for protocol classification and 60 for SoH regression, and most features were related to resistances, highly related features should be removed to reduce model complexity. To remove highly correlated features, the Pearson correlation coefficient between all generated features was compared, and if the absolute coefficient was higher than 0.75, all features except the feature having the highest correlation with the target value, SoC for protocol classification and SoH for SoH regression, were removed. After removing similar features and min-max scaling to features, the optimized feature number and ML model were selected by repeating 30 times of 7-fold cross-validation using Pycaret library in Python. 17 Based on selected features and the ML model, hyperparameter tuning was conducted for each model using Bayesian optimization. Model performances for each model were calculated by splitting the data into 15% for test data. For the classification model, accuracy was calculated to confirm performance of the model, and the equation was followed. $$\:Accuracy=\:\frac{Number\:of\:correct\:predictions}{Total\:number\:of\:predictions}$$ For the regression model, R 2 score, mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) were calculated to confirm the model performance, and the equations were followed. $$\:{R}^{2}\:score=1-\frac{\sum\:_{i=1}^{n}({{y}^{i}-{\widehat{y}}^{i})}^{2}}{\sum\:_{i=1}^{n}{({y}^{i}-\stackrel{-}{y})}^{2}}$$ $$\:MAE=\:\frac{\sum\:_{i=1}^{n}\left|{y}^{i}-{\widehat{y}}^{i}\right|}{n}$$ $$\:RMSE=\:\sqrt{\frac{\sum\:_{i=1}^{n}{({y}^{i}-{\widehat{y}}^{i})}^{2}}{n}}$$ $$\:MAPE=\frac{100}{n}\sum\:_{i=1}^{n}\left|\frac{{y}^{i}-{\widehat{y}}^{i}}{{y}^{i}}\right|$$ where \(\:{y}^{i}\) was true value, \(\:{\widehat{y}}^{i}\) was predicted value of i th data, \(\:\stackrel{-}{y}\) was mean of true values, and n was the amount of data. References Xu C et al (2022) Future greenhouse gas emissions of automotive lithium-ion battery cell production. Resour Conserv Recycl 187:106606. https://doi.org/10.1016/j.resconrec.2022.106606 . https://doi.org: Stroe DI, Schaltz E (2020) Lithium-Ion Battery State-of-Health Estimation Using the Incremental Capacity Analysis Technique. IEEE Trans Ind Appl 56:678–685. https://doi.org:10.1109/TIA.2019.2955396 Zhang S et al (2019) Synchronous estimation of state of health and remaining useful lifetime for lithium-ion battery using the incremental capacity and artificial neural networks. J Energy Storage 26:100951. https://doi.org:https://doi.org/10.1016/j.est.2019.100951 Park S et al (2023) Deep-learning based spatio-temporal generative model on assessing state-of-health for Li-ion batteries with partially-cycled profiles. Mater Horiz 10:1274–1281. https://doi.org:10.1039/D3MH00013C Zhu J et al (2022) Data-driven capacity estimation of commercial lithium-ion batteries from voltage relaxation. Nat Commun 13:2261. https://doi.org:10.1038/s41467-022-29837-w Tao S et al (2024) Generative learning assisted state-of-health estimation for sustainable battery recycling with random retirement conditions. Nat Commun 15:10154. https://doi.org:10.1038/s41467-024-54454-0 Kim D, Koo K, Jeong JJ, Goh T, Kim SW (2013) Second-Order Discrete-Time Sliding Mode Observer for State of Charge Determination Based on a Dynamic Resistance Li-Ion Battery Model. Energies 6:5538–5551 Lundberg SM, Lee S-I (2017) in Proceedings of the 31st International Conference on Neural Information Processing Systems 4768–4777Curran Associates Inc., Long Beach, California, USA Ovejas VJ, Cuadras A (2019) Effects of cycling on lithium-ion battery hysteresis and overvoltage. Sci Rep 9:14875. https://doi.org:10.1038/s41598-019-51474-5 Jang J et al (2025) Exploring the distinct effects of ionic and electronic conductivities of cathodes on the electrochemical performance of lithium-ion batteries. J Energy Storage 114:115935. https://doi.org/10.1016/j.est.2025.115935 . https://doi.org: Geslin A et al (2025) Dynamic cycling enhances battery lifetime. Nat Energy 10:172–180. https://doi.org:10.1038/s41560-024-01675-8 Hess A et al (2015) Determination of state of charge-dependent asymmetric Butler–Volmer kinetics for LixCoO2 electrode using GITT measurements. J Power Sources 299:156–161. https://doi.org/10.1016/j.jpowsour.2015.07.080 . https://doi.org: Chen Y et al (2022) Revealing the rate-limiting electrode of lithium batteries at high rates and mass loadings. Chem Eng J 450:138275. https://doi.org/10.1016/j.cej.2022.138275 . https://doi.org: Xu M, Reichman B, Wang X (2019) Modeling the effect of electrode thickness on the performance of lithium-ion batteries with experimental validation. Energy 186:115864. https://doi.org/10.1016/j.energy.2019.115864 . https://doi.org: Mishra AK, Shukla M (2025) Effect of Electrode Thickness and Operating Temperature on Electrochemical Performance of Li-Ion Batteries. Energy Storage 7:e70172. https://doi.org:https://doi.org/10.1002/est2.70172 Fan Y, Xiao F, Li C, Yang G, Tang X (2020) A novel deep learning framework for state of health estimation of lithium-ion battery. J Energy Storage 32:101741. https://doi.org:https://doi.org/10.1016/j.est.2020.101741 Ali M (2020) PyCaret: An open source, low-code machine learning library in Python , < https://pycaret.org/ Lain MJ, Brandon J, Kendrick E (2019) Design Strategies for High Power vs. High Energy Lithium Ion Cells Batteries 5:64 Stapf N et al (2025) Is Silicon Replaceable? A Physical, Chemical, and Electrochemical Analysis of Different Commercial Lithium-Ion Battery Cells. J Electrochem Soc 172. https://doi.org:10.1149/1945-7111/add112 Schmitt C, Gerle M, Kopljar D, Friedrich K (2023) Full Parameterization Study of a High-Energy and High-Power Li-Ion Cell for Physicochemical Models. J Electrochem Soc 170:070509. https://doi.org:10.1149/1945-7111/ace1a7 Jung T-J et al (2022) Statistical and computational analysis for state-of-health and heat generation behavior of long-term cycled LiNi0.8Co0.15Al0.05O2/Graphite cylindrical lithium-ion cells for energy storage applications. J Power Sources 529:231240. https://doi.org:https://doi.org/10.1016/j.jpowsour.2022.231240 Tables Table 1. Protocol 1 for SoH prediction Protocol 1 (Low SoC) Step Mode Voltage [V] (Cut-off C-rate for CV) Current [C-rate] Time [sec] 1 Charge (CC) 4.25 1.65 5 2 Rest - - 20 3 Discharge (CC/CV) 2.50 (0.02) 0.10 Compensation for capacity of step 1 (88 for CC-only case) 4 Rest - - 20 5 Charge (CC) 4.25 2.00 5 6 Rest - - 20 7 Discharge (CC/CV) 2.50 (0.02) 0.10 Compensation for capacity of step 5 (100 for CC-only case) 8 Rest - 20 2. Protocol 2 for SoH prediction Protocol 2 (Middle SoC) Step Mode Voltage [V] (Cut-off C-rate for CV) Current [C-rate] Time [sec] 1 Discharge (CC) 2.45 1.65 5 2 Rest - - 20 3 Charge (CC) 4.25 1.65 5 4 Rest - - 20 5 Discharge (CC) 2.45 2.00 5 6 Rest - - 20 7 Charge (CC) 4.25 2.00 5 8 Rest - - 20 Table 3. Protocol 3 for SoH prediction Protocol 3 (High SoC) Step Mode Voltage [V] (Cut-off C-rate for CV) Current [C-rate] Time [sec] 1 Discharge (CC) 2.45 1.65 5 2 Rest - - 20 3 Charge (CC/CV) 4.20 (0.02) 0.10 Compensation for capacity of step 1 (88 for CC-only case) 4 Rest - - 20 5 Discharge (CC) 2.45 2.00 5 6 Rest - - 20 7 Charge (CC/CV) 4.20 (0.02) 0.10 Compensation for capacity of step 5 (100 for CC-only case) 8 Rest - - 20 Table 4 . Cell specifications 19-21 Model Standard capacity Cathode Anode Cell 1 INR18650-29E 2.85 Ah NCA Graphite Cell 2 INR21700-50E 4.9 Ah NCA Graphite + Si Cell 3 INR21700-M50LT 4.8 Ah NCM Graphite Additional Declarations The authors declare no competing interests. Supplementary Files SoHpredictionsupportinginformation.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7524783","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":509560768,"identity":"ddb2037c-e624-4aa6-bd72-e62c07b26688","order_by":0,"name":"Minjeong Gong","email":"","orcid":"","institution":"Korea Advanced Institute of Science and Technology (KAIST)","correspondingAuthor":false,"prefix":"","firstName":"Minjeong","middleName":"","lastName":"Gong","suffix":""},{"id":509560769,"identity":"ebd186ba-f1e2-478a-a88d-52313730b6ea","order_by":1,"name":"Yoonjung Choi","email":"","orcid":"","institution":"Korea Advanced Institute of Science and Technology (KAIST)","correspondingAuthor":false,"prefix":"","firstName":"Yoonjung","middleName":"","lastName":"Choi","suffix":""},{"id":509560770,"identity":"776fb6ce-44fa-48c6-8705-cfb551ecf6bd","order_by":2,"name":"Han Mo Yang","email":"","orcid":"","institution":"SK ecoplant","correspondingAuthor":false,"prefix":"","firstName":"Han","middleName":"Mo","lastName":"Yang","suffix":""},{"id":509560771,"identity":"fab174fe-e4dc-42e8-8f79-061d1b0f0734","order_by":3,"name":"Sang Bok Ma","email":"","orcid":"","institution":"SK ecoplant","correspondingAuthor":false,"prefix":"","firstName":"Sang","middleName":"Bok","lastName":"Ma","suffix":""},{"id":509560772,"identity":"e2e1be10-5f18-4ecb-bdd9-1d260956918e","order_by":4,"name":"Dong-Hwa Seo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYDCCA0D84QADD5SbQJwWxhkgLWykaGHmAVlFtBa+84cPPrY5YycjP7+B8cMPhrR8glokb6QlG+fcSOYxOMbALNnDkGPZQEiLwQ0eM+mcD8w8BkCHSTMwVBgQtMXg/Plv0hYf6nnk2xiYfxOn5UAOmzTDjcM8DMcYgAyGHMJagH4xNuw5cxzol8Q2yx6DNMJagCH28MGPY9X28s2HD9/4UZFMWAsSYGwAupMUDaNgFIyCUTAKcAIARGg4ioYp5A0AAAAASUVORK5CYII=","orcid":"","institution":"Korea Advanced Institute of Science and Technology (KAIST)","correspondingAuthor":true,"prefix":"","firstName":"Dong-Hwa","middleName":"","lastName":"Seo","suffix":""}],"badges":[],"createdAt":"2025-09-03 08:49:42","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7524783/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7524783/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90891568,"identity":"1e510c42-15b9-472a-a36c-34196fc42922","added_by":"auto","created_at":"2025-09-09 11:11:41","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":80648,"visible":true,"origin":"","legend":"\u003cp\u003eOverall SoH prediction process for the spent batteries\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/303205331978f7ffe7677a84.jpg"},{"id":90891569,"identity":"74ce63fa-aaf0-449e-8888-ad602eba4a94","added_by":"auto","created_at":"2025-09-09 11:11:41","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":121056,"visible":true,"origin":"","legend":"\u003cp\u003eData collection and feature generation for protocol classification and SoH regression model (a) data collection method (b) SoH distribution (c) generated features for protocol classification model and (d) data collection method (e) SoH distribution (f) generated features for SoH regression model\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/5c105b82156eb6ccd41dc9dc.jpg"},{"id":90891576,"identity":"9bda59f7-1787-4bf6-85ce-15de257881c2","added_by":"auto","created_at":"2025-09-09 11:11:41","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":121264,"visible":true,"origin":"","legend":"\u003cp\u003eProtocol classification results (a) discharge voltage profile and protocol data corresponding to SoC at BoL (b) protocol classification model results (test data ratio = 15 %). The bubble sizes represent the number of data, and labels represent the number of misclassification cases for each train and test data (173 and 31 data for each SoC in train and test set). SHAP values for each feature (c) OCV, (d) R\u003csub\u003e0-0.1s\u003c/sub\u003e, (e) R\u003csub\u003e0.1-10s\u003c/sub\u003e and (f) skew (color bar: scaled feature value)\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/287699b42dd0d1a9fe4d58d5.jpg"},{"id":90894036,"identity":"d240d36a-6afa-41b1-bed5-0b8d056cbec3","added_by":"auto","created_at":"2025-09-09 11:27:41","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":184447,"visible":true,"origin":"","legend":"\u003cp\u003eSoH prediction results and analysis according to protocols (a) SoH prediction results (test data ratio = 15 %), (b) SoH prediction results from data measured protocol at untrained SoC, (c) SoH prediction results with misclassified protocol and SHAP values depending on the scaled top 3 features (d) 1.65C_dch_P_R\u003csub\u003e0-0.1s\u003c/sub\u003e, (e) 2.0C_ch_R_R\u003csub\u003e0-0.1s\u003c/sub\u003e and (f) 1.65C_dch_P_R\u003csub\u003e0.1-5s\u003c/sub\u003e. Color bars represent the SoH of each data.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/4da4ac23472cb830ed38e875.jpg"},{"id":90891575,"identity":"aecc68b6-f5a9-412d-bbea-8cc60333f5ef","added_by":"auto","created_at":"2025-09-09 11:11:41","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":106258,"visible":true,"origin":"","legend":"\u003cp\u003eApplication of SoH prediction process to the other cells (a) voltage profile, (b) dV/dQ during discharge at BoL, (c) relation between loading levels and cell capacity-volume ratio\u003csup\u003e18-20\u003c/sup\u003e, (d) protocol classification (The bubble sizes represent the number of data, labels represent the number of misclassification cases for each train and test data, and 18 and 15 data for each SoC in cell 2 and 3), (e) SoH regression prediction results of other cells (black edge symbols represent large error case due to different reaction mechanism), and (f) performance comparison according to the consumption time for SoH prediction\u003csup\u003e2-6,16\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/103ae7ea83fed479a664b058.jpg"},{"id":90897430,"identity":"b5fdd94d-984f-428d-a1bd-a4d1736fd461","added_by":"auto","created_at":"2025-09-09 11:43:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1417170,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/768ab545-754d-45a5-8176-32793e0b6792.pdf"},{"id":90891571,"identity":"dbb2a205-6c84-469c-b86b-4544317126c4","added_by":"auto","created_at":"2025-09-09 11:11:41","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2159660,"visible":true,"origin":"","legend":"","description":"","filename":"SoHpredictionsupportinginformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-7524783/v1/5d6ef6754722c28adf9fbd3d.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eFast and direct diagnosis of states of health for the spent batteries\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe demand for lithium-ion batteries is increasing continuously as electric vehicles (EVs) are commercialized to reduce greenhouse gas (GHG) emissions from internal combustion engines (ICEs). Thus, the number of spent batteries from EVs will also increase drastically in the near future. However, GHG emissions occur during the production of new batteries from material synthesis to battery manufacturing, although EVs can reduce GHG emissions by substituting ICEs during driving.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Thus, if the spent batteries can be used in energy storage systems (ESSs) or low-power EVs as a second life of batteries, it can effectively reduce GHG emissions from battery production.\u003c/p\u003e\u003cp\u003eThe states of the spent battery should be checked for evaluation of suitability for a second life. The important one among them is the state of health (SoH) that represents the remaining capacity relative to capacity at the beginning of life (BoL). To measure the SoH, it takes at least 6 hours per cell with a 0.33 C-rate for full charge and discharge. Thus, it takes considerable time to measure SoH for all the spent batteries. To quickly diagnose SoH of the spent batteries, researchers tried to predict SoH by using machine learning (ML), and there are two approaches to collect data for feature generation from galvanostatic measurements, such as voltage changes during constant current (CC) or rest time after current. The first approach was voltage data collection during CC charge, and it needed all or partial voltage profiles during charging to generate features for SoH prediction.\u003csup\u003e\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e To do this, it needs additional time and energy to set a fully discharged state or a specific voltage. Another approach was collecting voltage data at a specific state of charge (SoC). The voltage changes during rest after fully charging (100% SoC) or repeating current pulses and rest times at a specific SoC were used as features.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e In the latter case, it is important to set an accurate specific SoC due to various resistances depending on SoC.\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e Both approaches require additional pre-settings to get data for feature generation. On the other hand, Park \u003cem\u003eet al.\u003c/em\u003e reported SoH prediction without any pre-setting by regenerating the empty region in the entire charge and discharge.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e However, the performance was largely dependent on the amount of collected voltage profile, that if the time for voltage measurement was short, the performance of SoH prediction drastically decreased.\u003c/p\u003e\u003cp\u003eTo solve these problems, we suggest a SoH prediction process without any settings such as voltage or SoC. At this time, different protocols are applied to the spent batteries depending on the SoC near the upper or lower voltage cut-off, and it is determined based on voltages during the 1.0 C-rate discharge process for 10 seconds. Based on data collected from protocols containing various current pulses and rest times, numerous features are generated, and a trained ML model predicts the SoH of the spent cell. The overall process is conducted within 110 seconds (maximum 290 seconds), 10 seconds for protocol classification and 100 seconds (maximum 280 seconds) for protocol data, and it showed high performance for direct SoH prediction without any voltages or SoC settings.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n\u003ch2\u003eSoH prediction process\u003c/h2\u003e\n\u003cp\u003eThe overall process of SoH prediction for the spent batteries with unknown SoC was shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e., and there were two parts, protocol classification and SoH regression. Because voltage reached the cut-off voltage at high and low SoC when high current was applied, a different protocol was applied to get voltage data for SoH prediction. There were three types of protocol depending on the SoC (low, middle, and high), and voltage data measured from 1.0 C-rate discharge for 10 seconds were used for classification of protocol. Then, the SoH regression model predicted SoH based on the collected data applied a classified protocol containing several current pulses and rest times.\u003c/p\u003e\n\u003cp\u003eData was generated by using 18650 cylindrical cells (INR18650-29E) with 2.85 Ah of standard capacity. The 3 protocols were applied as shown in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cstrong\u003eSupplementary Fig.\u0026nbsp;3\u003c/strong\u003e. The number and boundary of protocols were determined based on the voltage profile, and the details will be discussed later. To train the protocol classification model, open-circuit voltage (OCV) and voltage data of 1.0 C-rate discharge for 10 seconds were collected with 5% SoC reduction from 100 to 0% SoC, and measured at 188 various SoH (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cstrong\u003e(a-b)\u003c/strong\u003e). For each collected voltage data, 8 features such as OCV, V\u003csub\u003e0s\u003c/sub\u003e V\u003csub\u003e10s\u003c/sub\u003e, R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;10s\u003c/sub\u003e, R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e, skew, and kurt were generated (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cstrong\u003e(c)\u003c/strong\u003e). On the other hand, to train the SoH prediction regression model, firstly, the discharge capacities of degraded cells with various cycle test conditions were measured with 0.5 C-rate charge and discharge to calculate SoH (\u003cstrong\u003eSupplementary Table\u0026nbsp;1\u003c/strong\u003e). Then, voltage data were collected at 5% SoC interval depending on the corresponding protocol that was determined based on the SoC (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cstrong\u003e(d)\u003c/strong\u003e). As shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cstrong\u003e(e)\u003c/strong\u003e, protocol data from 0 to 100% SoC were collected in 1043 various SoH. Due to getting data from various C-rate and current direction at each SoC and SoH, it was necessary to express their changes as features. Thus, features were generated to express voltage changes, kurtosis, and skewness for each C-rate (1.65 and 2.0C), pulse/rest (P/R), current direction (charge/discharge as ch/dch). Additionally, relations between C-rates or current directions, such as voltage difference after current pulse (hyst) and resistance ratio (R_ratio), were also used as features, and finally, a total 60 features were generated (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. \u003cstrong\u003e(f)\u003c/strong\u003e).\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003eProtocol classification\u003c/h3\u003e\n\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. \u003cstrong\u003e(a)\u003c/strong\u003e, the lower current (0.1 C) was applied in charge for higher 80% SoC and discharge for lower 10% SoC because these SoC regions were too close to cut-off voltage (\u003cstrong\u003eSupplementary Fig.\u0026nbsp;3\u003c/strong\u003e). Thus, it should be classified which protocol is applied to the spent cell, and the classification was conducted by using OCV, R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e, and skew as features that were selected after removing high correlation between features. Among various ML models, LGBMClassifier showed the highest accuracy as 0.9997 for train and 0.9985 for test data (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. \u003cstrong\u003e(b)\u003c/strong\u003e and \u003cstrong\u003eSupplementary Table\u0026nbsp;2\u003c/strong\u003e). At this time, there were misclassified cases at the near boundary between protocols with very low possibility, and the effects of protocol misclassification on SoH prediction will be discussed later.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe SHAP (Shapley Additive exPlanations) analysis was employed to understand how to classify protocols from 4 features, and can analyze the impact of each feature to predict the target value.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e The SHAP value means that the larger the SHAP value, the predicted value increases, and vice versa. In the case of a classification model, the SHAP value is calculated for each probability of classifying a certain class. Thus, the high SHAP value means the probability of classifying a certain class increases with a large impact.\u003c/p\u003e\n\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. \u003cstrong\u003e(c-f)\u003c/strong\u003e, OCV showed the highest impact for the protocol classification. OCV is directly related to the SoC, thus, low, middle, and high OCV contribute to increasing the probability of prediction with protocol 1, 2, and 3.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e represented the voltage difference between OCV and voltage at 0.1 seconds after the current pulse and was related to the ohmic resistance, containing electrical conductivity.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e Because the electrical conductivity of NCA and graphite at the low SoC was lower than the middle and high SoC, the high R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e contributed positively to protocol 1.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e And, R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e was directly related to the voltage difference during the current pulse for 10 seconds. Due to the large voltage drop at the low SoC, the high R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e positively contributed to the prediction. Oppositely, low R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e contribute positively to protocol 2 because the slope of voltage profile was less steep than the low SoC.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Similarly to R\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e, the high value of skew feature contributed positively to protocol 3, and in contrast, low skew contributed positively to protocol 1. Because voltage plateaus at high SoC and large voltage drops at low SoC, skewness at high and low SoC showed high and low values, and it affected the prediction.\u003c/p\u003e\n\u003ch3\u003eSoH regression\u003c/h3\u003e\n\u003cp\u003eTo predict SoH, the ML model trained on protocol data measured at all SoCs in 1043 SoH and generated 60 features as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. \u003cstrong\u003e(e-f)\u003c/strong\u003e. Among various ML models, RandomForest regressor showed the best performance to predict SoH (\u003cstrong\u003eSupplementary Table\u0026nbsp;3\u003c/strong\u003e). Model optimizations, such as feature selection and hyperparameter tuning, were conducted by using the RandomForest regressor, and finally, 12 features were selected (\u003cstrong\u003eSupplementary Table\u0026nbsp;4\u003c/strong\u003e). In Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. \u003cstrong\u003e(a)\u003c/strong\u003e, the optimized SoH prediction model showed high performance regardless of protocols with 0.998, 0.984% of R\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e score, 0.232, 0.850% of MAE, 0.414, 1.267% of RMSE, and 0.287, 1.055% of MAPE for the train and test data, respectively.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe SoH prediction model also showed high performance, with 2.235% of MAE, 2.741% of RMSE, and 2.799% of MAPE, from the data measured at the untrained SoC because SoH prediction model was trained on data measured at all SoC regions (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. \u003cstrong\u003e(b)\u003c/strong\u003e). In addition, the performances where data was measured from the misclassified protocol were also confirmed (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. \u003cstrong\u003e(c)\u003c/strong\u003e). The generated features were not similar to the train data due to different protocols, thus, RMSE, MAE, and MAPE slightly increased to 3.871% of MAE, 4.586% of RMSE, and 4.624% of MAPE. However, the accuracy for classification of protocol was very close to 1, thus, these cases will be rare, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. \u003cstrong\u003e(b)\u003c/strong\u003e. The resistance of cells varies depending on their SoC due to different ionic/electronic conductivity or reaction mechanisms.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e Thus, in most studies, ML models were trained on data measured at a specific SoC or certain voltage regions\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. On the other hand, our SoH prediction model was trained on features containing resistance measured in all SoCs. As a result, the SoH prediction model predicted with high accuracy, even though it was untrained on data such as SoC or protocol misclassification.\u003c/p\u003e\n\u003cp\u003eThe SHAP analysis was conducted to understand how SoH prediction model was trained, and 3 features with the highest mean absolute SHAP value were visualized in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. \u003cstrong\u003e(d-f)\u003c/strong\u003e. These features were related to resistance and highly correlated with SoH (\u003cstrong\u003eSupplementary table 4\u003c/strong\u003e). The SHAP values of these features decreased as feature values increased due to increasing resistances with cell degradation. These 3 features also varied depending on the SoC, similar to the previous study (\u003cstrong\u003eSupplementary Fig.\u0026nbsp;4\u003c/strong\u003e).\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e The 1.65C_dch_P_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, which was the highest correlation with SoH, was the most evenly distributed in all SoC ranges. However, the resistance at 100% and below 10% SoCs showed high resistance because resistances were calculated from the voltage difference, and large voltage drops were observed at those regions. Therefore, it should be compensated by using other features to increase accuracy at 100% and below 10% SoCs. First, 2.0C_ch_R_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e at 100% SoC showed much lower values compared to other SoCs because a constant voltage of 4.2 V was applied. Thus, the scaled 2.0C_ch_R_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e between 0 and 0.25 at protocol 3 were measured at 100% SoC, and the SHAP values showed different tendencies as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. \u003cstrong\u003e(e)\u003c/strong\u003e. On the other hand, 1.65C_dch_P_R\u003csub\u003e0.1\u0026minus;5s\u003c/sub\u003e were selected to increase accuracy below 10% SoCs. As shown in \u003cstrong\u003eSupplementary Fig.\u0026nbsp;4. (c)\u003c/strong\u003e, 1.65C_dch_P_R\u003csub\u003e0.1\u0026minus;5s\u003c/sub\u003e was observed in a wide range due to abrupt voltage drop at low SoCs. Thus, it could accurately predict SoH values, although higher 1.65C_dch_P_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e than other SoCs.\u003c/p\u003e\n\u003ch3\u003eExpandability of SoH prediction process\u003c/h3\u003e\n\u003cp\u003eTo confirm expandability, the suggested SoH prediction process was applied to the other cells. 2 types of 21700-sized cylindrical cells with different electrode chemistry and standard capacity were used (Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e), and their voltage profile and dV/dQ were shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cstrong\u003e(a, b)\u003c/strong\u003e. The voltage drops at the low SoC of cell 2 started at a higher SoC due to incorporation of Si in the anode. And the peak near SoC 80% in dV/dQ of cell 3 was shifted to slightly lower SoC due to different cathode chemistry.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cp\u003eBecause the values of generated features showed a large difference due to the different form factors of cell 2 and 3 compared to cell 1, it needs feature engineering for scaling from 21700 to 18650-sized cells to predict SoH by using SoH regression model trained on data from cell 1. Largely contributed features were related to resistance, such as dV\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e and dV\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e for protoco1 classification and 1.65C_dch_P_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, 2.0C_ch_R_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, and 1.65C_dch_P_R\u003csub\u003e0.1\u0026minus;5s\u003c/sub\u003e for SoH regression. Thus, we concentrated on the features related to the resistance for feature engineering, such as rescaling the resistance of different form factor cells to cell 1. There are numerous components affecting resistance, such as ionic/electronic conductivity, loading level, and electrode density. However, it was difficult to identify the exact information of the spent batteries, such as active material chemistry, loading level, and electrode density without additional analysis after cell disassembly. Thus, feature engineering without any detailed information was needed, and the ratio between cell capacity and external volume, such as 18650 and 21700 size, was used for feature engineering. As shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cstrong\u003e(c)\u003c/strong\u003e, the ratio of cell capacity and external volume showed a linear relationship with cathode and anode loading levels. When the loading level was higher, the resistance of the cell increased due to a longer diffusion path and higher tortuosity.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Thus, the features related to resistance were rescaled based on the ratio of cell capacity and external volume.\u003c/p\u003e\n\u003cp\u003eTo classify protocols for other cells, feature engineering was conducted on the features related to resistance using the ratio between cell capacity and external volume. Because the resistances were not directly related to the ratio between cell capacity and external volume, though it showed a linear relation with loading level, slope and intercept optimization were conducted on the features related to resistance, dV\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e and dV\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e. The optimized equations of each feature for rescaling from 21700 to 18650-sized cell were shown in \u003cstrong\u003eSupplementary Fig.\u0026nbsp;5\u003c/strong\u003e. After feature engineering to dV\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e and dV\u003csub\u003e0.1\u0026minus;10s\u003c/sub\u003e, it showed a similar tendency to cell 1. Thus, the accuracy also increased from 0.939 to 0.963 for cell 2, and the accuracy at low SoC was especially improved (\u003cstrong\u003eSupplementary Fig.\u0026nbsp;6\u003c/strong\u003e). However, different active materials chemistries caused accuracy reduction compared to cell 1, although feature engineering was performed. Thus, because these different chemistries were used in the cells and OCV, which largely dependent on chemistry, was the most effective feature for protocol classification, 10% SoC in cell 2 and 75% SoC in cell 3 were still misclassified (\u003cstrong\u003eSupplementary Fig.\u0026nbsp;7\u003c/strong\u003e). In case of cell 2, graphite with Si was used as an anode material, and it caused the voltage drop between 0% and 20% SoCs. In contrast, the NCM as a cathode material was used in cell 3 compared to NCA in cell 1. Thus, the peak position in dV/dQ of cell 3 at near 80% SoC was slightly lower than cell 1, causing misclassification at 75% SoC.\u003c/p\u003e\n\u003cp\u003eThen, protocol data were collected based on the SoC for SoH prediction, and SoH predictions for cell 2 and 3 were performed by using a SoH regression model trained on data from cell 1 data. And feature engineering was also conducted to feature related to resistance, similarly to protocol classification. Among various resistance-related features, the 3 features with the highest mean absolute SHAP value, 1.65C_dch_P_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, 2.0C_ch_R_R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e, and 1.65C_dch_P_R\u003csub\u003e0.1\u0026minus;5s\u003c/sub\u003e, were selected for feature engineering to evaluate the prediction performance for the other cells (\u003cstrong\u003eSupplementary Fig.\u0026nbsp;8\u0026ndash;10\u003c/strong\u003e). After feature engineering, the MAE of cell 2 and 3 were reduced from 4.230% and 4.768\u0026ndash;2.602% and 2.589%, respectively (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cstrong\u003e(e)\u003c/strong\u003e and \u003cstrong\u003eSupplementary Fig.\u0026nbsp;11\u003c/strong\u003e). Similar to the protocol classification results, large errors were observed at the different reaction occurred due to chemistry differences (black-edged marker in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cstrong\u003e(e)\u003c/strong\u003e).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, SoH of the spent battery was predicted within 110 seconds (maximum 290 seconds) with short current pulses and rests to classify protocols and predict SoH. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cb\u003e(f)\u003c/b\u003e, there were several studies to predict SoH of the spent batteries. Most studies concentrated on the data measured at specific voltage ranges, thus, voltage setting was necessary to predict SoH.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e Or it needed SoC setting to predict SoH because data collection and training model were conducted at a specific SoC in previous studies.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e On the other hand, it was reported that SoH was predicted by regeneration of the empty region in the entire voltage profile, thus, no settings were needed.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e However, the performance was critically affected by the amount of empty region. Thus, it took several hours to improve the performance. To address these problems, we suggested SoH prediction process regardless of SoC of the spent batteries without any SoC and voltage setting by training on all SoC range data to the regression model for SoH prediction. The suggested process could be used to predict SoH directly with high accuracy and very short consumption time for SoH diagnosis.\u003c/p\u003e\u003cp\u003eIn addition, other cells with different form factors can also be predicted through feature engineering. Feature engineering was conducted to only resistance-related features by using minimum cell information, such as the ratio between cell capacity and external volume that relates to the electrode loading level, and it effectively improved the performance of protocol classification and SoH regression.\u003c/p\u003e\u003cp\u003eSoH of the spent batteries for a second life can be determined within a short time through the suggested SoH prediction process with protocols and ML models, and energy and time to diagnose SoH for the spent battery will be saved. Furthermore, in this study, feature engineering was only conducted on the feature related to resistance, assuming electrode material chemistries were unknown. If it is possible to know the exact chemistries of the electrode material, feature engineering for other features can also be conducted with a suitable method to improve performance. The SoH of other form factors and chemistry cells can be predicted through feature engineering without additional data collection and training ML models, thus, energy and time are also saved to collect additional data of various cells.\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, 3 types of protocols containing various current pulses and rest time were applied, and ML models predicted a suitable protocol and SoH for the spent batteries with unknown SoC. The protocol classification model showed a high accuracy of 0.9985 for test data. SoH regression model predicted SoH based on the voltages measured through the corresponding protocol and showed a high R\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e score of 0.984 for the test data. The SoH prediction model also showed high performance in case of misclassification and untrained SoC by training on data in all SoCs. In addition, the performance of other cells with different form factors was improved through feature engineering to resistances-related features using the ratio between cell capacity and external volume, even though these cell data were not trained. The performance of SoH predictions will be improved through feature engineering reflecting the chemistry of electrode materials.\u003c/p\u003e"},{"header":"Method","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003eData generation\u003c/h2\u003e\u003cp\u003eCommercial cylindrical lithium-ion battery (LIB) cells, INR18650-29E (NCA/Graphite, Samsung SDI), were used for data generation. The chemistry and standard discharge capacity of the cell were summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e4\u003c/span\u003e. We used a 5 V and 12 A maximum cycler (BTS-4008-5V12A-S1, Neware) for the cell cycle test, protocol classification test, discharge capacity check, and three types of protocol tests. All the LIB cells were tested in a chamber maintained in room temperature (RT). Each LIB cell was cycled in the voltage range of 2.5\u0026ndash;4.2 V with different charge and discharge C-rate conditions that are summarized in \u003cb\u003eSupplementary Table\u0026nbsp;1\u003c/b\u003e: a constant current (CC) / constant voltage (CV) mode for charging step (end condition: 0.02 C-rate cut-off) and a CC mode for discharge step at RT. The sample size of each cycle test was 3 EA except for the test condition with 4.0 C-rate discharge, 6 EA.\u003c/p\u003e\u003cp\u003eThe protocol classification test, the discharge capacity check and protocol test were performed under the adjustment of cycle number to get data from evenly distributed SoH in SoH regions we interested in (\u003cb\u003eSupplementary Fig.\u0026nbsp;1\u003c/b\u003e). In protocol classification test, the voltage data during 1.0 C-rate discharge and rest time for 1 hour after discharge were collected at 5% depth of discharge (DoD) intervals from state of charge (SoC) 100\u0026ndash;0%.\u003c/p\u003e\u003cp\u003eThen, the discharge capacity was measured at 0.5 C-rate (CC mode) to check SoH, and the protocol tests were conducted, that divided by three different SoC ranges: 0, 5% SoC (Protocol 1, Low SoC), 10 to 75% SoC (Protocol 2, Middle SoC), and 80 to 100% SoC (Protocol 3, High SoC). In the protocol test, the cells were discharged at 1.0 C-rate, 5% DoD intervals from 100 to 0% SoC to set the specific SoC we were interested in. After setting SoC, they were conducted with protocol 3 to 1 with two different C-rates, 1.65 C-rate and 2.0 C-rate for 5 seconds and rest time for 20 seconds to collect the voltage data for SoH regression model. In protocol 1 and 3, we applied 0.1C-rate current in discharge and charge steps, respectively, because of the cut-off voltage condition. We summarized each protocol test condition according to three different SoC ranges in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e3\u003c/span\u003e (\u003cb\u003eSupplementary Fig.\u0026nbsp;3)\u003c/b\u003e.\u003c/p\u003e\u003cp\u003eTo confirm the performance of SoH prediction model at untrained SoCs, the protocol tests data were collected with three different initial SoC states, 98%, 97% and 96% SoC in various SoH states of 18650-29E cells. The cells were fully charged with 1.0 C-rate in CC/CV mode (end condition: 0.02 C-rate cut-off) and then discharged with 1.0 C-rate under the end conditions of 2%, 3% and 4% DoD based on the discharge capacity we checked, respectively. There were the same as the trained SoC cases in the whole protocol test conditions according to the SoC ranges and 5% DoD intervals except for the end SoC states, 3%, 2% and 1% for 98%, 97% and 96% SoC in initial states, respectively. For the misclassified protocol test, we artificially applied the upper and lower levels of the protocol condition in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e3\u003c/span\u003e after setting the particular SoC in the cells with various SoH states. We divided four cases according to SoC values, 5%, 14%,75% and 84% SoC, respectively. In case 1 (5% SoC), the cells were fully charged and then discharged with 1.0 C-rate under the end-condition of 95% DoD based on the discharge capacity we checked. After setting 5% SoC, we conducted the test with protocol 2 (middle SoC), not protocol 1, to check the performance of SoH prediction model. In case 2 (14% SoC), case 3 (75% SoC) and case 4 (84% SoC), the SoC setting method was the same as case 1, and the protocol test were conducted in each case: case 2 (14% SoC) with protocol 1 (low SoC), case 3 (75% SoC) with protocol 3 (high SoC), and case 4 (84% SoC) with protocol 2 (middle SoC), respectively.\u003c/p\u003e\u003cp\u003eFor verification of SoH prediction process suggested, two commercial 21700-sized LIB cells, INR21700-50E (NCA/Graphite\u0026thinsp;+\u0026thinsp;Si, Samsung SDI) and INR21700-M50LT (NMC/Graphite, LG Chem), were used to obtain cell data. The chemistry and standard discharge capacity of these cells are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e4\u003c/span\u003e. In the case of 21700-sized LIB cells, we conducted the cycle test with 0.5 C-rate charge (CC/CV mode) / 1.0 C-rate discharge (CC mode) at RT. The sample size of each cycle test was 3 EA. We collected 1.0 C-rate discharge data for protocol classification, discharge capacity, and protocol data every 100 cycles under the same condition and environments as INR18650-29E.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eData preprocessing\u003c/h2\u003e\u003cp\u003eFrom the data collected at 0.5 C-rate for protocol classification and SoH regression, the SoC and SoH were calculated using the following equations.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:SoC\\:\\left[\\%\\right]=\\:\\frac{Remaining\\:capacity}{Fully\\:discharge\\:capacity}\\times\\:100$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:SoH\\:\\left[\\%\\right]\\:=\\:\\frac{Discharge\\:capacity}{Discharge\\:capacity\\:of\\:BoL}\\times\\:100$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThen, features were calculated from voltage data for protocol classification and SoH regression after discharge capacity measurement. In case of the protocol classification model, 3 types of values were generated, such as actual voltages, resistances, and distribution of voltage during discharge. The actual voltages were open circuit voltage (OCV), V\u003csub\u003e0.1s\u003c/sub\u003e, and V\u003csub\u003e10s\u003c/sub\u003e (V\u003csub\u003ex s\u003c/sub\u003e is voltage after x seconds during 1.0 C-rate discharge). At this point, quasi-OCV was used as OCV, which was the end voltage after 1-hour rest time. Resistances were calculated as follows:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{R}_{x-y\\:s}=\\:\\frac{\\left|{V}_{x\\:s}-{V}_{y\\:s}\\right|}{current}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere V\u003csub\u003ex s\u003c/sub\u003e was the voltage at x seconds during 1.0 C-rate discharge, and V\u003csub\u003e0s\u003c/sub\u003e was OCV. The distribution of voltage during discharge was characterized by skewness and kurtosis of voltages for a 10-second 1.0 C-rate discharge. In the case of SoH regression model, the generated values were similar to the protocol classification model, such as actual voltages, resistances, and distributions of voltage during pulse or rest. Additionally, the relations between C-rate or current direction were also expressed as values because various current pulses and rests were contained in protocols. V\u003csub\u003e0.1s\u003c/sub\u003e and V\u003csub\u003e5s\u003c/sub\u003e as actual voltages, R\u003csub\u003e0\u0026thinsp;\u0026minus;\u0026thinsp;0.1s\u003c/sub\u003e and R\u003csub\u003e0.1\u0026minus;5s\u003c/sub\u003e as resistances, and skewness and kurtosis as distributions of voltage were generated to express each current pulse or rest. And end voltage difference at each current pulse as hysteresis (hyst), resistance ratio between charge and discharge at the same C-rate as R_ratio, and resistance ratio between 1.65 and 2.0 C-rate were generated as features to express the relation between C-rate and current direction. The features for SoH regression model were numerous, thus, features were labeled as (C-rate)_(charge/discharge)_(pulse/rest)_(value) for each pulse and rest.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eMachine Learning\u003c/h2\u003e\u003cp\u003eTo classify protocol and predict SoH by using machine learning (ML), feature selection and model optimization, such as model selection and hyperparameter tuning, should be conducted. The same procedures were conducted for protocol classification and SoH regression model. Because there were numerous generated features, 8 features for protocol classification and 60 for SoH regression, and most features were related to resistances, highly related features should be removed to reduce model complexity. To remove highly correlated features, the Pearson correlation coefficient between all generated features was compared, and if the absolute coefficient was higher than 0.75, all features except the feature having the highest correlation with the target value, SoC for protocol classification and SoH for SoH regression, were removed.\u003c/p\u003e\u003cp\u003eAfter removing similar features and min-max scaling to features, the optimized feature number and ML model were selected by repeating 30 times of 7-fold cross-validation using Pycaret library in Python.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e Based on selected features and the ML model, hyperparameter tuning was conducted for each model using Bayesian optimization. Model performances for each model were calculated by splitting the data into 15% for test data. For the classification model, accuracy was calculated to confirm performance of the model, and the equation was followed.\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\:\\frac{Number\\:of\\:correct\\:predictions}{Total\\:number\\:of\\:predictions}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFor the regression model, R\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e score, mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) were calculated to confirm the model performance, and the equations were followed.\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:{R}^{2}\\:score=1-\\frac{\\sum\\:_{i=1}^{n}({{y}^{i}-{\\widehat{y}}^{i})}^{2}}{\\sum\\:_{i=1}^{n}{({y}^{i}-\\stackrel{-}{y})}^{2}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:MAE=\\:\\frac{\\sum\\:_{i=1}^{n}\\left|{y}^{i}-{\\widehat{y}}^{i}\\right|}{n}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equg\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equg\" name=\"EquationSource\"\u003e\n$$\\:RMSE=\\:\\sqrt{\\frac{\\sum\\:_{i=1}^{n}{({y}^{i}-{\\widehat{y}}^{i})}^{2}}{n}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equh\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equh\" name=\"EquationSource\"\u003e\n$$\\:MAPE=\\frac{100}{n}\\sum\\:_{i=1}^{n}\\left|\\frac{{y}^{i}-{\\widehat{y}}^{i}}{{y}^{i}}\\right|$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{y}^{i}\\)\u003c/span\u003e\u003c/span\u003e was true value, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\widehat{y}}^{i}\\)\u003c/span\u003e\u003c/span\u003e was predicted value of i\u003csup\u003eth\u003c/sup\u003e data, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{-}{y}\\)\u003c/span\u003e\u003c/span\u003e was mean of true values, and n was the amount of data.\u003c/p\u003e\u003c/div\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eXu C et al (2022) Future greenhouse gas emissions of automotive lithium-ion battery cell production. Resour Conserv Recycl 187:106606. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.resconrec.2022.106606\u003c/span\u003e\u003cspan address=\"10.1016/j.resconrec.2022.106606\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eStroe DI, Schaltz E (2020) Lithium-Ion Battery State-of-Health Estimation Using the Incremental Capacity Analysis Technique. IEEE Trans Ind Appl 56:678\u0026ndash;685. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1109/TIA.2019.2955396\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1109/TIA.2019.2955396\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang S et al (2019) Synchronous estimation of state of health and remaining useful lifetime for lithium-ion battery using the incremental capacity and artificial neural networks. J Energy Storage 26:100951. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:https://doi.org/10.1016/j.est.2019.100951\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1016/j.est.2019.100951\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark S et al (2023) Deep-learning based spatio-temporal generative model on assessing state-of-health for Li-ion batteries with partially-cycled profiles. Mater Horiz 10:1274\u0026ndash;1281. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1039/D3MH00013C\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1039/D3MH00013C\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu J et al (2022) Data-driven capacity estimation of commercial lithium-ion batteries from voltage relaxation. Nat Commun 13:2261. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41467-022-29837-w\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41467-022-29837-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTao S et al (2024) Generative learning assisted state-of-health estimation for sustainable battery recycling with random retirement conditions. Nat Commun 15:10154. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41467-024-54454-0\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41467-024-54454-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim D, Koo K, Jeong JJ, Goh T, Kim SW (2013) Second-Order Discrete-Time Sliding Mode Observer for State of Charge Determination Based on a Dynamic Resistance Li-Ion Battery Model. Energies 6:5538\u0026ndash;5551\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg SM, Lee S-I (2017) in \u003cem\u003eProceedings of the 31st International Conference on Neural Information Processing Systems\u003c/em\u003e 4768\u0026ndash;4777Curran Associates Inc., Long Beach, California, USA\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOvejas VJ, Cuadras A (2019) Effects of cycling on lithium-ion battery hysteresis and overvoltage. Sci Rep 9:14875. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41598-019-51474-5\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41598-019-51474-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJang J et al (2025) Exploring the distinct effects of ionic and electronic conductivities of cathodes on the electrochemical performance of lithium-ion batteries. J Energy Storage 114:115935. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.est.2025.115935\u003c/span\u003e\u003cspan address=\"10.1016/j.est.2025.115935\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGeslin A et al (2025) Dynamic cycling enhances battery lifetime. Nat Energy 10:172\u0026ndash;180. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41560-024-01675-8\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41560-024-01675-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHess A et al (2015) Determination of state of charge-dependent asymmetric Butler\u0026ndash;Volmer kinetics for LixCoO2 electrode using GITT measurements. J Power Sources 299:156\u0026ndash;161. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jpowsour.2015.07.080\u003c/span\u003e\u003cspan address=\"10.1016/j.jpowsour.2015.07.080\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen Y et al (2022) Revealing the rate-limiting electrode of lithium batteries at high rates and mass loadings. Chem Eng J 450:138275. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cej.2022.138275\u003c/span\u003e\u003cspan address=\"10.1016/j.cej.2022.138275\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu M, Reichman B, Wang X (2019) Modeling the effect of electrode thickness on the performance of lithium-ion batteries with experimental validation. Energy 186:115864. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.energy.2019.115864\u003c/span\u003e\u003cspan address=\"10.1016/j.energy.2019.115864\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMishra AK, Shukla M (2025) Effect of Electrode Thickness and Operating Temperature on Electrochemical Performance of Li-Ion Batteries. Energy Storage 7:e70172. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:https://doi.org/10.1002/est2.70172\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1002/est2.70172\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFan Y, Xiao F, Li C, Yang G, Tang X (2020) A novel deep learning framework for state of health estimation of lithium-ion battery. J Energy Storage 32:101741. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:https://doi.org/10.1016/j.est.2020.101741\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1016/j.est.2020.101741\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAli M (2020) \u003cem\u003ePyCaret: An open source, low-code machine learning library in Python\u003c/em\u003e, \u0026lt;\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pycaret.org/\u003c/span\u003e\u003cspan address=\"https://pycaret.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLain MJ, Brandon J, Kendrick E (2019) Design Strategies for High Power vs. High Energy Lithium Ion Cells Batteries 5:64\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eStapf N et al (2025) Is Silicon Replaceable? A Physical, Chemical, and Electrochemical Analysis of Different Commercial Lithium-Ion Battery Cells. J Electrochem Soc 172. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1149/1945-7111/add112\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1149/1945-7111/add112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchmitt C, Gerle M, Kopljar D, Friedrich K (2023) Full Parameterization Study of a High-Energy and High-Power Li-Ion Cell for Physicochemical Models. J Electrochem Soc 170:070509. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1149/1945-7111/ace1a7\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1149/1945-7111/ace1a7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJung T-J et al (2022) Statistical and computational analysis for state-of-health and heat generation behavior of long-term cycled LiNi0.8Co0.15Al0.05O2/Graphite cylindrical lithium-ion cells for energy storage applications. J Power Sources 529:231240. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:https://doi.org/10.1016/j.jpowsour.2022.231240\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1016/j.jpowsour.2022.231240\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1.\u0026nbsp;Protocol 1 for SoH prediction\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"599\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003e\n \u003cp\u003eProtocol 1 (Low SoC)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eStep\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVoltage [V]\u003c/p\u003e\n \u003cp\u003e(Cut-off C-rate for CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCurrent\u003c/p\u003e\n \u003cp\u003e[C-rate]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTime [sec]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC/CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.50\u003c/p\u003e\n \u003cp\u003e(0.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCompensation for capacity of step 1\u003c/p\u003e\n \u003cp\u003e(88 for CC-only case)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC/CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.50\u003c/p\u003e\n \u003cp\u003e(0.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCompensation for capacity of step 5\u003c/p\u003e\n \u003cp\u003e(100 for CC-only case)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e2.\u0026nbsp;Protocol 2 for SoH prediction\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"599\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003e\n \u003cp\u003eProtocol 2 (Middle SoC)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eStep\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVoltage [V]\u003c/p\u003e\n \u003cp\u003e(Cut-off C-rate for CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCurrent\u003c/p\u003e\n \u003cp\u003e[C-rate]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTime [sec]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 3.\u0026nbsp;Protocol 3 for SoH prediction\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"599\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003e\n \u003cp\u003eProtocol 3 (High SoC)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eStep\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVoltage [V]\u003c/p\u003e\n \u003cp\u003e(Cut-off C-rate for CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCurrent\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e[C-rate]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTime [sec]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC/CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.20\u003c/p\u003e\n \u003cp\u003e(0.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCompensation for capacity of step 1\u003c/p\u003e\n \u003cp\u003e(88 for CC-only case)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDischarge (CC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCharge (CC/CV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.20\u003c/p\u003e\n \u003cp\u003e(0.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCompensation for capacity of step 5\u003c/p\u003e\n \u003cp\u003e(100 for CC-only case)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u003c/strong\u003e. Cell specifications\u003csup\u003e19-21\u003c/sup\u003e\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 9.2165%;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 28.1107%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eStandard capacity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eCathode\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAnode\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 9.2165%;\"\u003e\n \u003cp\u003eCell 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 28.1107%;\"\u003e\n \u003cp\u003eINR18650-29E\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.85 Ah\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eNCA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGraphite\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 9.2165%;\"\u003e\n \u003cp\u003eCell 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 28.1107%;\"\u003e\n \u003cp\u003eINR21700-50E\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.9 Ah\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eNCA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGraphite + Si\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 9.2165%;\"\u003e\n \u003cp\u003eCell 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 28.1107%;\"\u003e\n \u003cp\u003eINR21700-M50LT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.8 Ah\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eNCM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGraphite\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Korea Advanced Institute of Science and Technology","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"SoH, Machine learning, spent battery, lihtium-ion battery","lastPublishedDoi":"10.21203/rs.3.rs-7524783/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7524783/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eFast diagnosis of state of health (SoH) is essential for a second life of the spent batteries due to drastically increasing demands for lithium-ion batteries. However, additional voltage or state of charge (SoC) settings were needed to predict SoH. Here, we suggest a SoH prediction process without any additional settings within 110 seconds (maximum 290 seconds) using machine learning, which consists of 2 steps: protocol classification and SoH regression. Data were generated from 18650-sized cells. There are 3 protocols depending on SoC due to cut-off voltage, and open-circuit voltage and 1.0 C-rate discharge voltage for 10 seconds were collected at every 5% SoC in 188 SoH. The protocol classification model showed 0.9985 of accuracy. Then, protocol data, with current pulses and rest times, corresponding to SoC were collected at every 5% SoC in 1043 SoH, and 12 features were selected from 60 features. The SoH regression model showed 0.850% mean absolute error (MAE), and performance in cases of untrained SoC and misclassification was also superior because the model was trained on all SoC data. In addition, the suggested SoH prediction process can apply to untrained cells, such as different form factors. The performances of 21700-sized cells were improved to 2.600% MAE through feature engineering to the resistance-related features using the ratio between cell capacity and external volume. This paper highlights a fast and direct SoH prediction process without any settings, and feature engineering of untrained cell data to save time and energy for diagnosis of spent batteries.\u003c/p\u003e","manuscriptTitle":"Fast and direct diagnosis of states of health for the spent batteries","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-09 11:11:36","doi":"10.21203/rs.3.rs-7524783/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"605f89c4-1866-459f-adf6-ffaa2ed2e82a","owner":[],"postedDate":"September 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54120792,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-09-09T11:11:37+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-09 11:11:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7524783","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7524783","identity":"rs-7524783","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}