Accurate Chemistry Identification of Lithium-Ion Batteries Based on Temperature Dynamics with Machine Learning

Abstract

Lithium-ion batteries (LIBs) are widely used in diverse applications, ranging from portable ones to stationary ones. The appropriate handling of the immense amount of spent batteries has, therefore, become significant. Whether recycled or repurposed for second-life applications, knowing their chemistry type can lead to higher efficiency. In this paper, we propose a novel machine learning-based approach for accurate chemistry identification of the electrode materials in LIBs based on their temperature dynamics under constant current cycling using gated recurrent unit (GRU) networks. Three different chemistry types, namely lithium nickel cobalt aluminium oxide cathode with silicon-doped graphite anode (NCA-GS), nickel cobalt aluminium oxide cathode with graphite anode (NCA-G), and lithium nickel manganese cobalt oxide cathode with graphite anode (NMC-G), were examined under four conditions, 0.2 C charge, 0.2 C discharge, 1 C charge, and 1 C discharge. Experimental results showed that the unique characteristics in the surface temperature measurement during the full charge or discharge of the different chemistry types can accurately carry out the classification task in both experimental setups, where the model is trained on data under different cycling conditions separately and jointly. Furthermore, experimental results show that the proposed approach for chemistry type identification based on temperature dynamics appears to be more universal than voltage characteristics. As the proposed approach has proven to be efficient in the chemistry identification of the electrode materials LIBs in most cases, we believe it can greatly benefit the recycling and second-life application of spent LIBs in real-life applications.

1. Introduction

As the use of lithium-ion batteries grows across many consumer electronics, mobile, and stationary applications, the need to consider them for second-life applications or recycling them is becoming more expedient due to the environmental impact of disposing of waste LIBs and the increasing cost of lithium and other constituent materials in the market. Among the recycling methods, direct recycling adopts the direct relithiation of the cathode materials for reuse in manufacturing new cells [1]. However, this method is very sensitive to cathode materials because the direct recycling method needs to be tailored to specific cathode materials to obtain an acceptable product quality [2]. The other recycling methods, pyrometallurgy and hydrometallurgy, can recover valuable metals such as nickel and cobalt while additional steps are needed to adequately recover lithium in a form that can be directly used in building new cells. Yet, knowing the chemistry of the spent LIB helps identify the potential metals that can be recovered. The information can guide designing complex recycling processes to recycle all such cells or adopting a dedicated recycling process after sorting the cells according to the chemistry [3]. In repurposing for second-life applications, it is recommended to combine cells with similar capacity and similar chemistry in secondary applications [4]. The reason is that different electrode chemistries have different performance patterns and age differently, such that combining them may lead to safety risks and poor performance [5].

Unfortunately, the identification of the chemistry type of the LIB remains a daunting challenge. Existing methods such as energy dispersive X-ray spectrometry (EDS), X-ray photoelectron spectroscopy (XPS), Raman spectroscopy, and X-ray diffraction (XRD) are effective in the identification of the electrode material of the cell [6,7,8,9]. However, they are not fit for commercial cells, except the cells are opened and samples taken for testing. Moreover, these are expensive technologies and require specialised knowledge in using and interpreting them. Having a method to identify the chemistry without opening the commercial LIB will significantly reduce cost and time in second-life applications and recycling ventures.

Many different LIBs exist in the market, and in many situations, the constituent materials are not in the public domain because they are manufacturers’ company secrets. Examples of some commercial cathode materials include lithium nickel cobalt aluminium oxide (NCA), lithium cobalt oxide (LCO), lithium nickel manganese cobalt oxide (NMC), lithium manganese oxide (LMO), and lithium iron phosphate (LFP) [10,11,12,13]. Graphite is commonly used for the anode, although other variants, such as silicon-doped graphite, lithium metal and lithium titanate oxide, exist [12]. Although LFP as a cathode with graphite anode can easily be identified from the nominal voltage of 3.3 V on the label of the cell or the datasheet, the other layered-oxide cathodes above with graphite anode are challenging to recognise from the datasheet as their nominal voltage ranges between 3.6–3.7 V depending on the chemistry, manufacturer and model [14]. Typically, the nominal voltage is determined by the cathode and anode used in the cell. However, with advances in LIB research, some manufacturers also include some additives in electrodes and electrolytes that help improve electrical and ionic conductivity, charge transfer rates, and thermal stability and, in turn, impact the nominal voltage of their cells [15]. As a result, identifying the chemistry of LIB from the operating voltage is difficult.

Just as state estimations of LIBs are mainly obtained from the measurement of voltage, current, and temperature [16], they also contain helpful features that can be used to infer the chemistry type of the LIB. For example, the open circuit voltage (OCV) curve, which provides a relationship between the equilibrium voltage of the cell and the state of charge, shows a unique pattern based on the cell electrodes [17,18]. The OCV obtained from partial charge and discharge has been used to train a machine learning algorithm to identify the NMC and LFP cathode [19]. However, carrying out OCV measurements takes time because reaching the equilibrium voltage can take several hours. Since the temperature of the surface of the cell can be easily measured during operation and with the growing interest in the investigation of the temperature behaviour of LIB, especially in differential thermal voltammetry as a means of estimating SOH and studying the degradation modes in LIB, there is the possibility of useful pattern that can be used in chemistry identification. We hereby explore the equations that govern the heat generation in LIB.

Ignoring the negligible heat of mixing and the heat due to side reactions in normal operating LIB, total heat generation, ${\dot{Q}}_{\mathrm{total}}$ can be expressed as a sum of the joule heating, ${\dot{Q}}_{\mathrm{joule}}$ which is the generated heat due to current flowing through the cell resistance and the entropic heat, ${\dot{Q}}_{\mathrm{entropic}}$ which is either generated or extracted as a result of the change in the entropy of the electrodes materials as the lithium ions intercalate and deintercalate [20,21,22,23].

$${\dot{Q}}_{\mathrm{total}}\approx {\dot{Q}}_{\mathrm{joule}}+{\dot{Q}}_{\mathrm{entropic}}$$

(1)

and in terms of the open circuit voltage $V_{\mathrm{OCV}}$, terminal voltage V, cell internal temperature T, and current I:

$${\dot{Q}}_{\mathrm{total}}=I\left(V-V_{\mathrm{OCV}}\right)+IT{\displaystyle \frac{\partial V_{\mathrm{OCV}}}{\partial T}}$$

(2)

The joule heating, $I\left(V-V_{\mathrm{OCV}}\right)$ is influenced by the ohmic and polarisation resistance. Although temperature affects internal ohmic resistance, resistance due to polarisation is greatly influenced by SOC, temperature, and current during charge and discharge [24]. The joule heating is always exothermic. The $V_{\mathrm{OCV}}$, on the other hand, depends on the state of charge and shows a characteristic behaviour that is specific to the electrodes in the cell [25]. In addition, the entropic heat too is dependent on current, temperature and the entropy coefficient, ${\displaystyle \frac{\partial V_{\mathrm{OCV}}}{\partial T}}$, which is a function of the state of charge. Thus, over the full range of charge and discharge, a characteristic entropy heat is generated, which shows up in the thermal behaviour of the cell. In addition, the chemical composition of the anode and cathode affects the rate of reversible heat generated in the LIB, which varies with the state of charge (SOC) [22,26,27,28]. The entropic heat can either be exothermic or endothermic depending on the SOC and whether during charge or discharge.

The effect of the reversible heat is being explored in this work to identify the chemistry of the cell as it impacts the temperature of the surface of the cell during the charge and discharge phase. Unfortunately, the temperature at the cell’s surface is affected by the ambient temperature depending on the insulation and the thermal properties of the various components that make up the cell. These are the porous electrodes, binders, separators, current collectors, and electrolytes, each with its characteristic heat capacity and thermal conductivity. Various manufacturers select these materials depending on their design requirements. Steinhardt et al. analysed the result of several investigations on the thermal properties of the individual components and the lumped parameters of LIBs by various authors using different chemistry of LIB. They showed that the specific heat capacities varied based on the cell format, the ambient temperature of measurement, cell capacity and the SOC. It was found that the median is 912 J/Kg/K for cylindrical, 1168 J/Kg/K for pouch and 1041 J/Kg/K for prismatic and significant variation of 1314–795 J/Kg/K was observed in the heat capacities of the cylindrical cells [29]. To mitigate the effect of the difference in temperature between the cell surface and the ambient temperature on the identification of the chemistry of the cell electrode, very low thermal conductivity styrofoam (polystyrene foam), typically in the range of 0.03 to 0.04 W/m/K will be very beneficial [30]. Further to the use of good insulation, the impact of the various heat capacities of the same chemistry cells from different sources can be greatly reduced by the normalisation of the temperature measurement.

Thus, the chemistry of the LIB may be identified from the cell’s surface temperature, which indicates how much heat is generated based on the cell electrodes. This behaviour is mainly non-linear, and complex measurements and equations may be analysed to provide a good result. This is a typical classification problem that a machine learning algorithm would generally solve if supplied with enough data. In fact, machine learning approaches have been applied to the problem of chemistry type determination of batteries in some precedent works already. In ref. [31], the authors proposed a framework to determine the chemical composition of batteries using specially designed preprocessing procedures and two different classifiers, namely an artificial neural network and a classification tree. The proposed framework used voltage and current data that were obtained when the cells were connected to different loads and can accurately classify the five types of batteries under testing, including lithium nickel cobalt aluminium oxide, lithium iron phosphate, nickel–metal hydride and lithium titanate oxide. In ref. [32], the authors proposed a federated machine learning approach for battery recycling purposes, where low cathode sorting errors are achieved utilising features extracted from the end-of-life cycle among five different cathode materials, such as the peak intensity of the ${\displaystyle \frac{dV}{dQ}}$ curve during charging or discharging, the skewness statistics of the voltage, the kurtosis statistics of the capacity, and so on. In ref. [33], the authors proposed a supervised machine learning framework for accurately classifying different lithium–sulfur battery electrolytes. Despite the diverse implementation of classification algorithms, the ingenious design of input features has been necessary, which poses extra challenges to the problem. Our approach of simply using the surface temperature measurement in full charge or discharge phase does not require complicated feature extraction analysis.

Therefore, we propose a simple machine learning approach for determining the chemistry type of lithium-ion batteries solely using the temperature profiles during constant current cycling. The proposed approach takes full charge and discharge phase at different C-rates as input and can accurately classify the chemistry types given enough training samples. In this paper, three different lithium-ion cells, namely the NMC cathode/graphite anode, the NCA cathode/graphite anode and the NCA cathode/silicon-doped graphite anode, were aged at 1 C charge and discharge cycles with a check-up test performed at 0.2 C charge and discharge to obtain ageing data. Section 2 describes the experimental setup, while Section 3 discusses the machine learning methodology. Section 4 contains the result and discussion, and the conclusion is in Section 5.

2. Experimental Setup

The experimental setup consists of three cells for each of the following chemistries: lithium nickel cobalt aluminium oxide cathode with silicon-doped graphite anode (NCA-GS), nickel cobalt aluminium oxide cathode with graphite anode (NCA-G) and lithium nickel manganese cobalt oxide cathode with graphite anode (NMC-G) as shown in

Table 1. Test cells.

Cell Chemistry	Specification	Capacity (mAh)	Model Number	Manufacturer
NCA-GS	3 to 4.2 V	3120	US18650VTC6	Sony
NCA-G	3 to 4.3 V	3250	NCR-18650B	Panasonic
NMC-G	2.5 to 4.2 V	4800	INR21700 M50LT	LG

. NCA-G and NCS-GS are in 18650 format, while the NMC-G is in 21700 format.

A PT100 temperature sensor was placed directly mid-way on the curved surface of the cylindrical cells, and they were also insulated with 1.5 cm thick styrofoam on the top, bottom and around the curved surface. The setup for the 21700 cell format is shown in Figure 1. The insulation is used to limit the impact of the ambient temperature on the cells’ temperature profile during charge and discharge operation. Small holes were made at the top and bottom insulation cover for the sensor and the electrical connections. This set-up was placed inside the Memmert ICP110 temperature chamber, which provided the constant ambient temperature of 25 °C for the insulated cells, and an Arbin LBT21084UC battery tester was used for charging and discharging the cells.

The fast ageing cycle consists of constant-current (CC) charge at 1 C and corresponding CC discharge at 1 C without rest time. After every 25 cycles of 1 C charge and discharge, the cells were charged with constant-current constant-voltage charge (CCCV) at 0.2 C to full charge at 4.2 V high cut-off voltage and low cut-off current of 0.05C. Thereafter, a rest time of 5 h was provided before a CC discharge at 0.2 C was applied to the full discharge at the low cutoff voltage of 2.5 V. Another rest time of 5 h was provided before carrying out a CCCV charge at 0.2 C which provided the 0.2 C charge profile that was used in this work. The 0.2 C discharge cycle was used to check the state of health (SOH) of the cell as it ages, and the temperature measurements obtained during the charge and discharge at 0.2 C and 1 C were used in the machine learning algorithm. The ageing cycles at 1 C and the checkup cycles at 0.2 C were repeated until the cell reached about 75% SOH.

As part of the experiment, we applied galvanostatic intermittent titration technique (GITT) after every 50 ageing cycles. In this case, we discharge the cells after full charge with 0.2 C pulse for 2.4 min and a resting pulse for 6 min. The pulsed-discharge and rest sequence was repeated till full discharge. The result was used to find the internal resistance of the cells as a function of SOC with age. The internal resistance at each SOC was calculated by finding the difference between the voltage after rest time and the voltage at the instance of applying the discharge current pulse. The discharge capacity obtained during each pulse is used the determine the SOC and the equation for the approximate internal resistance which includes the ohmic and polarisation resistance in 1 s sample time is given below

$$R\approx {\displaystyle \frac{(V_{\mathrm{i}}-V_{\mathrm{rest}})}{I}}$$

(3)

3. Methodology

3.1. Gated Recurrent Unit

Gated recurrent units (GRUs) [34] are a type of refined recurrent neural networks (RNNs) for time series processing. Although the development of long short-term memory (LSTM) [35] partially solved the vanishing gradient problem of traditional RNNs by introducing the gate mechanism [36] and LSTM has proven capable in various time series processing tasks [37,38], the training of LSTM suffers from the excessive number of parameters. Based on this, GRUs were introduced with fewer gates to facilitate easy training of the network with fewer samples required [39,40]. Consequently, we use GRU as the basic unit for feature extraction in our work. Figure 2 shows the internal structure of GRUs.

The mapping equations are as follows:

$${\mathit{z}}_{\mathit{t}}=\sigma ({\mathit{W}}_{\mathit{z}}{\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{z}}{\mathit{h}}_{\mathit{t}-\mathit{1}}+{\mathit{b}}_{\mathit{z}})$$

(4)

$${\mathit{r}}_{\mathit{t}}=\sigma ({\mathit{W}}_{\mathit{r}}{\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{r}}{\mathit{h}}_{\mathit{t}-\mathit{1}}+{\mathit{b}}_{\mathit{r}})$$

(5)

$${\tilde{\mathit{h}}}_{\mathit{t}}=tanh({\mathit{W}}_{\mathit{h}}{\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{h}}({\mathit{r}}_{\mathit{t}}\odot {\mathit{h}}_{\mathit{t}-\mathit{1}})+{\mathit{b}}_{\mathit{h}})$$

(6)

$${\mathit{h}}_{\mathit{t}}=(1-{\mathit{z}}_{\mathit{t}})\odot {\mathit{h}}_{\mathit{t}-\mathit{1}}+{\mathit{z}}_{\mathit{t}}\odot {\tilde{\mathit{h}}}_{\mathit{t}}$$

(7)

where ${\mathit{x}}_{\mathit{t}}$ is the input vector, ${\mathit{h}}_{\mathit{t}}$ is the hidden state, ${\mathit{z}}_{\mathit{t}}$ is the update gate vector, ${\mathit{r}}_{\mathit{t}}$ is the reset gate vector, and ${\tilde{\mathit{h}}}_{\mathit{t}}$ is the candidate activation vector, at time step t, respectively. $\mathit{W}$ and $\mathit{U}$ are the corresponding weight matrices for each mapping equation. The update gate determines how much previous and current information should be retained in the memory with Equation (4), and the reset gate determines which past information should be forgotten with Equation (5). In Equation (6), the candidate activation vector ${\tilde{\mathit{h}}}_{\mathit{t}}$ is selected based on the retained past information and the current information. Finally, as shown in Equation (7), the final output vector ${\mathit{h}}_{\mathit{t}}$ is formed by combining the weighted old hidden state and the weighted new candidate vector.

3.2. Model Workflow

Figure 3 shows the workflow of the GRU model for battery chemistry type classification. The measured temperature profiles are normalised using min-max normalisation after the constant bias of the initial temperature reading is removed. This time series of temperature differences after normalisation is the only input of the GRU models. In the GRU model, temporal features are extracted by the GRU, where the feature vector at the last time step, considered to conclude the information of the whole sequence, is fed into the subsequent fully connected (FC) layers for the calculation of the logits, namely the unnormalised scores of the possibility of each battery type. The softmax function is then used to transform the logits into normalised possibilities. The corresponding battery type is determined based on the highest normalised or unnormalised score.

Table 2. Architecture of the GRU model.

Layer	Input Size	Output Size
GRU	1	64
FC	64	32
ReLU	-	-
FC	32	3

shows the architecture of the GRU model, where ReLU stands for the activation function rectified linear unit. We use one GRU layer for temporal feature extraction and two FC layers to map the feature vector into the vector of logits. In each experiment, the respective dataset is divided randomly into the training set, validation set, and test set with a ratio of 6:2:2. The training set is used for the updating of the model parameters. During the training process, the performance of the model is constantly validated on the validation set, which is then used as a metric for the fine-tuning of the hyperparameters. In our work, the most important hyperparameter, the learning rate, is fine-tuned using the automatic hyperparameter fine-tuning framework Optuna [41]. To ensure comparability, the same data division is used for the joint training experiment as the separate ones. The models’ performance on the test set is presented in the end.

4. Results and Discussion

4.1. Experimental Results

The first three charge and discharge cycles using 1 C of NCA-G are shown in Figure 4. In the 1 C cycle, there was no rest time between the charge and discharge phase, so the first charge cycle caused a rise in cell surface temperature from the ambient of 25 °C to 48 °C, 44 °C, and 39 °C in the NCA-G, NMC-G, and NCA-GS, respectively. The choice of no rest time between the charge and discharge was part of the lab ageing test. It is also useful to investigate whether regular cycles without rest time or steady-thermal state can also be used to infer the cell’s chemistry.

The test schedule of a single 0.2 C charge and discharge cycle for one of the NCA-G cells is shown in Figure 5. Due to the low current, the entropic heat is also more pronounced, as indicated by the characteristic temperature.

Figure 6 shows the SOH of the cells during 0.2 C discharge. This shows the variability of the dataset used in training the machine learning model as the cells had a slightly different number of cycles at 1 C and 0.2 C based on their unique cell difference even among the same chemistry.

The plot of the internal resistance as a function of SOC for a representative cell for each chemistry is shown in Figure 7. Clearly, it shows an increase in the internal resistance with age, thus affecting the shape of the temperature profile. The internal resistance curve is bow-shaped, as higher resistances are observed at the lower end and higher end of the SOC. Due to the linear proportionality of the rate of joule heating to the resistance, the rate of joule heating will also be bow-shaped. This would always lead to a time-delayed temperature increase in all cell chemistries during charging or discharging. Therefore, there is no likelihood of a drop in temperature measurement due to only joule heating, as it does not have any cooling effect and is always exothermic. The result of this is the elevated cell temperature in the first 1 C charge phase, as previously shown in Figure 4, which signifies that the rate of joule heating (W) is quite significant at the higher current of 1 C when compared with the 0.2 C charge due to the square effect of the current on the rate of joule heat generation. However, at elevated temperatures in the subsequent cycles, the cell’s internal resistance is lower than its value at room temperature, so the entropic heat’s cooling effect results in the temperature drop midway. The reversible heat is linearly proportional to the current and leads to the characteristic temperature profile due to different chemistries. This result corresponds to other works on joule and entropic heating in LIBs [42,43].

The surface temperature profile as the new cells aged to about 75% SOH of the various chemistries at 0.2 C charge, 0.2 C discharge, 1 C charge, and 1 C discharge are, respectively, shown in Figure 8, Figure 9, Figure 10 and Figure 11. The plot shows the difference between the cell surface temperature and the initial surface temperature at the start of charge or discharge. The plots are from a single cell, which is representative of similar cells of the same chemistry. In the 0.2 C cycles, the initial temperature is the ambient reference temperature of 25 °C, which had a sufficient rest phase of 5 h. A unique temperature profile can be observed on each chemistry at the beginning-of-life (BOL) for the 0.2 C current rate. The distinct pattern is more evident during the charge than during the discharge phase. However, the curves become smoother with age due to an increase in the resistance and a decrease in the capacity, so distinct patterns become blurred. This means that the characterisation will be less reliable when the cells are aged further.

In the 1 C cycles, the initial temperature is quite elevated, as previously mentioned; due to the higher charge 1 C, the temperature profile transits from a wave-like shape to a bow shape with age in the NCA-GS, whereas it transits from a bow shape to a wavelike shape in the NCA-G. NMC-G stays bow-like over the entire cell’s lifespan. In the discharge at 1 C, the NCA-G had a smooth bow shape throughout the age of the cell, while the NCA-GS and NMC-G had a rough bow shape that remained unique throughout their age. Within the limit of SOH between 100% and 75%, the joule heating was not too high to neglect the effect of the entropic heat, as we observe that the temperature is not a smooth bow but has some undulation that differentiates the cells. So, the contribution of entropic heat is still distinguishable. In addition, the overlap in the internal resistance of the NMC-G and the NCA-GS also shows that the distinguishing factor in the temperature profile is not their internal resistance but the entropic heat.

Each time-series data for every charge and discharge phase when using 1 C and 0.2 C and their corresponding electrode chemistry was used in the machine learning algorithm. For example, one of the NMC-G cells had 19 time-series data when discharging at 0.2 C and 19 time-series data when charging at 0.2 C from BOL up to approximately 75% SOH. Also, the 1 C charge and discharge were up to 451 time-series data. The number of the time-series data for each cell in the various measurement schemes is shown in

Table 3. Number of time series data of the temperature measurement.

Chemistry	Cell	0.2 C Charge	0.2 C Discharge	1 C Charge	1 C Discharge
NMC-G	D58	41	41	1001	1001
NMC-G	D59	19	19	451	451
NMC-G	D60	24	24	576	576
NCA-GS	N31	33	33	801	801
NCA-GS	N32	33	33	801	801
NCA-GS	N33	32	32	776	776
NCA-G	N51	6	6	126	126
NCA-G	N52	6	6	126	126
NCA-G	N53	7	6	151	151
Total		201	200	4809	4809
Training	60%	120	120	2885	2885
Validation	20%	40	40	962	962
Testing	20%	41	40	962	962

. The time-series data for each test category, namely 0.2 C charge, 0.2 C discharge, 1 C charge and 1 C discharge, irrespective of the chemistry, cell name, cycle number or SOH, were then randomly selected for training, validation and testing in the ratio 60:20:20. For example, for the 0.2 C charge, 120 time-series data were randomly selected for training, 40 time-series data for validation, and 41 time-series data for testing.

As shown in the above plots, the time-series data were preprocessed by normalising the temperature difference for the full charge data and the full discharge data so that the range of the normalised temperature is between 0 and 1. This is shown in Equation (8):

$${\mathit{T}}_{\mathrm{norm}}={\displaystyle \frac{(\mathit{T}-T_0)-min\{\mathit{T}-T_0\}}{max\{\mathit{T}-T_0\}-min\{\mathit{T}-T_0\}}}$$

(8)

where ${\mathit{T}}_{\mathrm{norm}}$ is the normalised time-series temperature array, $T_0$ is the initial temperature at the beginning of the charge or the discharge and $\mathit{T}$ is the actual time-series temperature measurement array. The result of this normalisation is shown in Figure 12 for three cells of different chemistry in the BOL during 0.2 C charge, 1 C charge, 0.2 C discharge, and 1 C discharge. The figure is representative of the various chemistries. These normalised data were used to train the GRU model in solving the classification problem.

4.2. Chemistry Type Classification with Machine Learning

To better analyse the performance of the machine learning approach on the classification of the chemistry type of lithium-ion batteries, two distinct experiments are carried out, namely separately trained cycling conditions and jointly trained cycling conditions. Under separately trained cycling conditions, four GRU models are trained and tested separately for the four presented cycling conditions, namely 0.2 C charge, 0.2 C discharge, 1 C charge, and 1 C discharge. Under jointly trained cycling conditions, one GRU model is trained with the mixed data of all four cycling conditions but tested individually with the data of each condition. In the case of separately trained cycling conditions, we train and test the four models for each condition to study the correlation between the distinguishable temperature dynamics and the cycling conditions, based on which detailed analyses and interpretations are carried out. In the case of jointly trained cycling conditions, we trained and tested one jointly trained model to check the proposed approach’s applicability in general cases where the cycling conditions do not have to be restricted. In addition, we compare the identification outcome based on temperature dynamics with voltage dynamics. Four evaluation metrics are used in this work, including accuracy, precision (Pr), recall (Re), and F1-score:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$$

(9)

$$Pr={\displaystyle \frac{TP}{TP+FP}}$$

(10)

$$Re={\displaystyle \frac{TP}{TP+FN}}$$

(11)

$$F1=2{\displaystyle \frac{Pr·Re}{Pr+Re}}$$

(12)

where TP, TN, FP, and FN stand for true positive, true negative, false positive, and false negative, respectively. Accuracy reflects the overall classification performance of the model, precision reflects the model’s ability not to identify a negative sample as positive, recall reflects the model’s ability to find all positive samples, and F1-score is the harmonic mean value of the precision and recall.

4.2.1. Performance Under Separately Trained Cycling Conditions

In this section, results from the separately trained cycling conditions are shown. For each of the four cycling conditions, one GRU model is trained and tested separately on the respective dataset. This experiment is aimed at determining the cycling condition with the strongest features that can be used for cell chemistry identification. Figure 13 and

Table 4. Comparison of the model’s performance trained and tested under different cycling conditions.

Condition	Accuracy [%]	Chemistry	Pr [%]	Re [%]	F1 [%]
0.2 C Charge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0
0.2 C Discharge	97.5	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	94.1	97.0
		NMC-G	95.0	100.0	97.4
1 C Charge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0
1 C Discharge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0

show the classification performance of the GRU model trained and tested separately under different cycling conditions. The confusion matrices visualise the true label of each sample and the category it is classified into intuitively. The number of test samples is simply the sum of all entries of each confusion matrix. For example, for the 0.2 C charge case (Figure 13a), there are 5 NCA-G test samples, 19 NCA-GS test samples, and 17 NMC-G test samples, yielding a total number of testing samples of 41. Each row represents the samples of the respective actual class, while each column represents the samples of the respective predicted class. The correctly categorised samples are, therefore, shown in the diagonal entries.

As shown in the confusion matrices and box plots in Figure 13, the proposed GRU model performs in most cases impeccably. Under the 0.2 C charge condition, as shown in Figure 13a,b, the model is able to identify all chemistry types with confidence close to 100%, which is the case for the cycling condition of 1 C discharge as well. Under the cycling condition of 1 C charge, the model has also achieved 100% accuracy, but its confidence in the classification of NCA-G samples becomes more scattered, with the lowest one being around 60%. The only sample that is wrongly identified by the model appears under the cycling condition of 0.2 C discharge, where an NCA-GS sample is classified as NMC-G by the model. Under this cycling condition, the overall accuracy of the classification becomes 97.5%. In this case, the recall rate of NCA-GS samples is 94.1% and the precision of NMC-G samples is 95.0%, resulting in the F1-scores of 97.0% and 97.4% for NCA-GS and NMC-G, respectively. The overall interval of classification confidence on NMC-G samples moves slightly downwards as well, but still almost all above 80%.

Because of the test schedule design, cycling data from 0.2 C is very limited compared with 1 C, especially for the NCA-G cells. However, the lack of data does not cause significant errors in such cases, which proves the fact that the temperature profiles during constant current cycling have unique characteristics with respect to the chemistry types of the cells, and thus can be used for accurate identification of cell types. The high confidence levels of classification under all four cycling conditions further support this statement.

4.2.2. Performance Under Jointly Trained Cycling Conditions

In this section, results from the jointly trained cycling conditions are shown. One GRU model is trained with the mixed data of all four cycling conditions and tested separately on the respective dataset. This experiment is aimed at exploring the possibility of using the temperature profiles under arbitrary cycling conditions for cell chemistry identification. Figure 14 and

Table 5. Comparison of the model’s performance trained with mixed data and tested under different cycling conditions.

Condition	Accuracy [%]	Chemistry	Pr [%]	Re [%]	F1 [%]
0.2 C Charge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0
0.2 C Discharge	97.5	NCA-G	80.0	100.0	88.9
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	94.7	97.3
1 C Charge	99.9	NCA-G	100.0	98.9	99.4
		NCA-GS	100.0	100.0	100.0
		NMC-G	99.7	100.0	99.9
1 C Discharge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0

show the classification performance of the GRU model under jointly trained cycling conditions.

As shown in Figure 14a,b, the GRU model trained with mixed data is able to achieve perfect accuracy and extremely high confidence in classification with minimal variance on the test data from the 0.2 C charge condition, where the accuracy, precision, recall, and F1-score are all 100%, which is the same case for the model’s performance under the 1 C discharge condition. Under the 0.2 C discharge condition, the model wrongly identifies one NMC-G sample as NCA-G, where its confidence in the correct class, namely NMC-G, is around 25%, resulting in an overall accuracy of 97.5%. Except for this one outlier, the classification confidence is all close to 100%. Higher uncertainties are shown under the 1 C charge condition, where the distribution of the model’s classification confidence on NCA-G and NMC-G samples becomes more scattered. In this case, one NCA-G sample is wrongly identified as an NMC-G sample.

The test results show that training the model with unbalanced data under different cycling conditions does not jeopardise its performance on the different samples. In fact, compared with being separately trained with the data from the respective condition, the model is able to achieve comparably impeccable performance on the cell type identification task when trained with mixed data of all cycling conditions. This shows that it is highly realistic to train one universal model with different kinds of cycling data and use it for the task of chemistry type identification under different cycling conditions.

Figure 15 shows the jointly trained model’s confidence on each sample’s real class with respect to the SOH. For most samples with a variety of different SOHs, the classification confidence is close to 100%, which proves that the proposed approach performs ideally on a wide range of SOHs. Around four outliers of NCA-G samples can be spotted out of a total number of samples of 176. In particular, three out of the four outliers have in fact been correctly identified, as they still have a confidence over 50%, leaving only one wrongly identified sample. These outliers do not show a particular pattern with regard to SOH, which in fact lie across different aging states. The same goes for NMC-G samples, where four out of 819 samples can be considered outliers, which also do not show any particular pattern with regard to SOH. Combining the fact that the rest of the samples all have a confidence of almost 100% and the outliers scatter in different ranges of SOHs, it does not really indicate that there are any specific SOH ranges where the model does not work correctly.

4.2.3. Comparison with Identification Using Voltage Profiles

Voltage measurement is considered essential and simple to implement for battery cycling. Therefore, we would like to explore the possibility of cell chemistry type classification based on voltage dynamics for comparison with using temperature dynamics. For this purpose, we trained four GRU models with the same architecture as before but using voltage profiles during the four cycling conditions, respectively.

Table 6. Performance of the model trained with voltage profiles under different cycling conditions.

Condition	Accuracy [%]	Chemistry	Pr [%]	Re [%]	F1 [%]
0.2 C Charge	46.3	NCA-G	0.0	0.0	0.0
		NCA-GS	46.3	100.0	63.3
		NMC-G	0.0	0.0	0.0
0.2 C Discharge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0
1 C Charge	99.7	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	99.4	99.7
		NMC-G	99.2	100.0	99.6
1 C Discharge	100.0	NCA-G	100.0	100.0	100.0
		NCA-GS	100.0	100.0	100.0
		NMC-G	100.0	100.0	100.0

shows the performance of the respective models and Figure A1 in Appendix A shows the confusion matrices. For 1 C charge and discharge, the model could identify the cell chemistry accurately based on voltage dynamics as well, with an accuracy of 99.7% and 100.0%, respectively, just as with temperature dynamics. However, we observe that it becomes significantly harder for the model to converge during training under 0.2 C charge and discharge conditions, which could be caused by the limited amount of cycling data under such conditions. Therefore, the configurations of the early stopping technique have to be loosened greatly to allow for the slow learning process of the model. After elaborate fine-tuning on the two 0.2 C cycling conditions, the model trained with voltage profiles could reach 100% accuracy under the 0.2 C discharge condition as well. However, the model is proven unable to identify the cell chemistry types well from the voltage characteristics under the 0.2 C charge condition, where none of the NCA-G and NMC-G samples could be correctly identified, resulting in an overall accuracy of 46.3%. We suspect that the CV phase has a slight effect on the result since the voltage is the same and flat for all chemistries at the CV phase, and so there is no characteristic feature for differentiation, while the temperature drop during CV is different for the various chemistries, and therefore the temperature profile is better for identification.

Not surprisingly, the voltage dynamics also contain unique characteristics that can be used for chemistry type identification in most cycling cases, except for 0.2 C charge, which is not a problem when using temperature dynamics. Despite the fact that accurate voltage measurement is generally easier to implement, we believe that temperature dynamics provide more unique information originating from the different cell chemistries and thus appear to be more universal for the task of cell type identification.

5. Conclusions

The recycling of lithium-ion batteries is of great importance due to their dominance in the market shares. To conduct efficient and economical recycling, knowing the chemistry of the spent LIBs is very beneficial, as this information can guide the recycling processes. At the same time, an accurate identification of the chemistry types of the spent LIBs can lead to a better repurposing of LIBs for second-life applications as well. In this paper, we propose a novel machine learning-based approach for accurate chemistry identification of LIBs based on their temperature dynamics under constant current cycling using GRU networks. Three different chemistry types, namely NCA-GS, NCA-G, and NMC-G, were examined under four cycling conditions, including 0.2 C charge, 0.2 C discharge, 1 C charge, and 1 C discharge. In the case of separately trained cycling conditions, the model was able to achieve the accuracy of 100.0%, 97.5%, 100.0%, and 100.0% under the four cycling conditions, respectively, which showcases the fact that the unique characteristics in the temperature profiles of the different chemistry types can indeed be used to carry out the classification task accurately. In the case of jointly trained cycling conditions, the model was able to perform perfectly in all four cycling conditions with the accuracy of 100.0%, 97.5%, 99.9%, and 100.0%, which indicates that it is highly realistic to train one universal model with different kinds of cycling data and use it for the task of chemistry type identification under different cycling conditions. In addition, test results also show that temperature dynamics appear to be more reliable than voltage dynamics on the task with regard to the classification. We believe that the proposed approach has proven to be efficient for the chemistry identification of LIBs in most cases, and thus could benefit the recycling and second-life application of spent LIBs in real-life applications. This method is particularly useful in cases where the voltage limits from the datasheet are not sufficient to identify the cell. Another advantage of this method is that the temperature measurement during the fast 1 C charge or discharge of about an hour can be used to identify the electrode chemistry.

While the result of this approach of determining the cell’s chemistry is promising, there is still the challenge of training the model with cycling data of several cells from different manufacturers, of different capacities and form factors such as prismatic, cylindrical and pouch cells. These may have different joule heating contributions compared to the entropy that may influence the temperature profiles. The impact of the various cell thermal parameters and the ambient temperature, which was ignored due to the insulation in this work, may become significant in real-life applications where carrying out such insulation may be a complex activity depending on the cell format, cell pack format and battery system setup. The classification may also become a challenge for cells with a significant age of SOH below 75% where the irreversible heat can be so high that the temperature response may become less distinctive. Thus, more ageing data are required. These factors present an opportunity for future research to expand the training data and use them to determine the chemistry of the lithium-ion cell.