Artificial intelligence-based predictive reference model for lithium iron phosphate battery cell aging analysis
Raja
Yahmadi
1
Kais
Brik
1,2✉
Emailkais.brik@yahoo.fr
1
A
A
Laboratory Materials, Measurements and Applications
INSAT, University of Carthage Tunisia
2
Higher Institute of Multimedia Arts of Manouba
University of Manouba
Tunisia
Raja Yahmadi 1, Kais Brik 1,2
1
Laboratory Materials, Measurements and Applications, INSAT, University of Carthage Tunisia.
2
Higher Institute of Multimedia Arts of Manouba, University of Manouba, Tunisia
Corresponding author’s e-mail: kais.brik@yahoo.fr
Abstract
This study introduces an Artificial intelligence (AI) approach to model the discharge voltage characteristics of a new Lithium-Iron Phosphate (LFP) battery cell under different operating conditions and to use it as a reference for healthy assessment. Experimental voltage-State Of Charge (SOC) data were obtained from a new cell at three temperatures (0°C,25°C, and 45°C) and for several discharge currents. In order to predict the appropriate discharge voltage behavior under any operating conditions, a Gaussian Process Regression (GPR) model was trained using temperature, discharge current, and SOC as input variables. The trained model provides a continuous voltage reference under any realistic combination of temperature and current. Based on this reference, a diagnostic system was developed to compare the measured discharge voltage of cycled cells with the reference voltage of a new cell under the same conditions. The deviation between the predicted and measured voltages enables the estimation of State of Health (SOH) and allows assessing whether a manufactured cell exhibits early degradation. This approach provides a fast and efficient solution for cell quality assessment and early detection of abnormal degradation. The results demonstrate that the proposed AI based reference model enables reliable SOH evaluation, offering strong potential for industrial diagnostic applications and manufacturing quality control.
Keywords:
LFP cell
artificial intelligence
Gaussian Process Regression (GPR)
sate of health prediction
diagnostic system
A
1. Introduction
Lithium-Iron Phosphate (LFP) batteries are widely adopted in electric vehicles, renewable energy storage, and industrial applications due to their long cycle life, robust thermal stability, and safe electrochemical behavior [1–5]. Despite these advantages, LFP cells experience aging mechanisms such as loss of active material and increase internal resistance that affect voltage response, reduce usable capacity, and limit power capability [6–8]. Accurate characterization of battery aging is therefore critical for ensuring reliability, performance, and safety in energy storage systems [9]. However, their electrochemical behavior especially the voltage response during discharge is highly dependent on operating conditions such as temperature, discharge rate, and SOC, making health assessment under varying conditions particularly challenging [2, 10–11].
In industrial cell manufacturing, battery validation programs typically combine cycling and calendar aging tests conducted at several controlled temperatures, generally ranging from 0° to 45°C. These protocols are designed to activate and accelerate different degradation mechanisms, enabling comprehensive characterization of cell aging behavior across a wide operational range. Despite this diversity of test conditions, SOH assessment is conventionally performed under standardized reference conditions, most commonly at 25°C and a C/3 discharge current rate, to ensure consistent and comparable of diagnostic results.
As a result, whenever SOH evaluation is required during ongoing cycling or calendar aging tests conducted at non reference temperatures and current rate, the cell must first be brought back to 25°C and subjected to a dedicated low-discharge current rate diagnostic protocol. This requirement leads to test interruptions, introduces additional test downtime, increases operational complexity, and reduces overall test throughput, particularly in large-scale validation programs.
Similar constraints are encountered in filed applications, such as renewable energy storage systems, where batteries operate continuously under highly variable thermal, electrical, and load conditions. In such systems, direct SOH monitoring is particularly challenging, as conventional diagnostic procedures would require disconnecting battery strings and enforcing standardized conditions (25°C and discharge current C/3), which is neither technically practical nor operationally feasible in real-world installations.
Conventional diagnostic methods for lithium-ion batteries are generally based on performance comparisons between an aged cell and a new reference cell tested under identical operating conditions. Previous studies have extensively investigated battery health estimation techniques. For exemple, Che et al. [10] reviewed the mechanisms and methods for battery health prognostics, highlighting established diagnostic techniques, but these approaches often fail under variable temperature and current conditions, limiting their applicability in real-world scenarios. Similarly, Severson et al. [13] demonstrated that early-cycle voltage features can predict cycle life with high accuracy; however, their method requires complete early cycling datasets and is not designed for single-discharge prediction. More recently, Attia et al. [14] propose data-efficient machine learning frameworks combined with Bayesian optimization to improve prediction accuracy, but the approach still depends on long experimental sequences, making it impractical for rapid or field level diagnostics.
Furthermore, strict reproduction of identical temperature, current rate, and soc conditions is difficult to achieve in practice and even minor deviations can significantly affect diagnostic indicators, leading to inconsistent SOH estimation. Repeated execution of controlled diagnostic tests therefore results in increased testing time, higher operational costs, and limited scalability. These constraints significantly reduce the suitability of conventional SOH assessment methods for in-situ, real-time, or condition-aware monitoring.
To overcome these limitations, there is a strong industrial and application-driven need for advanced diagnostic approaches that can estimate battery SOH directly under ongoing cycling, calendar aging, or operational conditions, without interrupting test protocols or modifying operating conditions. This requirement motivates the development of the proposed AI-based diagnostic framework and its associated modular, condition-aware interface designed for real-time monitoring and analysis.
In recent years, machine learning has become a powerful tool for analyzing complex electrochemical behaviors and predicting critical battery parameters from limited datasets [15, 16]. For example, Liu et al [15] applied deep learning for fast real-world diagnostics, but performance drops under variable conditions. Li et al [17] reviewed machine learning applications in battery development, highlighting the potential of AI-based diagnostics while noting the lack of methods robust to varying SOC, temperature, and current rate. Motivated by these limitations, this study introduces an AI-based approach capable of predicting the discharge voltage behavior of a new cell under arbitrary operating conditions. By learning the nonlinear interactions between current rate, temperature, and SOC, the model generates a reliable reference voltage response without the need to replicate standardized experiments. This approach enables fast, cost-effective, and condition-independent health diagnosis, offering a more practical and scalable alternative to conventional technique.
GPR has emerged as a powerful technique for modeling nonlinear battery behaviors due to its flexibility, uncertainty estimation and ability to generalize from limited datasets, making it ideal for both laboratory and industrial environments. In this study, GPR is used to develop a predictive reference model for the voltage-SOC behavior of a new LFP cell under various temperatures and discharge currents. This model serves as a virtual healthy baseline for comparing with cycled cells, enabling accurate health assessment.
2. Methodology
This study presents an AI-based model for battery health assessment using new LFP cells as a reference for comparison comprises three main steps. In the first stage, new cells are experimentally characterized to obtain their discharge voltage profiles across different temperature and discharge currents. These measurements provide a high-quality dataset for model training.
In the second stage, a GPR model is developed to capture the nonlinear relationships among voltage, temperature, discharge current, and SOC. The model is trained using the experimental dataset and then employed to predict the expected voltage response of a new cell under any operating condition.
Finally, the predicted discharge voltage provides a reference for evaluating the health of tested cells. Deviations between the measured and predicted voltage are quantified to assess aging and capacity loss. By eliminating the need to reproduce the exact test conditions of a new reference cell, this approach enables fast, cost-effective, and condition-independent diagnosis.
The overall workflow of the proposed methodology is illustrated in Fig. 1, which summarizes the experimental characterization, model training, and diagnostic evaluation process.
Fig. 1
AI based Approach for battery health assessment using new LFP cells as a reference for comparison
3. Experimental Characterization of new LFP Cells
A commercial LFP cell was experimentally characterized under controlled laboratory conditions to construct a comprehensive voltage-SOC dataset. Three representative temperatures were selected (0°C, 25°C, and 45°C) to investigate the temperature dependence of the cell’s electrochemical response. For each temperature, multiple constant-current (CC) discharge tests were performed across a range of currents representative of typical operating conditions. During each test, the cell voltage was continuously recorded as a function of SOC using high-resolution data acquisition to capture subtle variations in the voltage response.
This experimental protocol resulted in a multi-operating-condition dataset capturing the combined influence of temperature and discharge current on the voltage-SOC behavior. Such a dataset allows the analysis of nonlinear behaviors, including voltage drop at high currents and the shift of plateau regions with temperature. The evolution of the discharge voltage under varying currents is depicted in Fig. 2A for 0°C, Fig. 2B for 25°C, and Fig. 2C for 45°C, illustrating the strong dependence of the LFP cell’s voltage response to operating conditions throughout the SOC range.
Fig. 2
Discharge Voltage Vs SOC of new LFP cells under varying currents and temperatures
At 0°C, all discharge curves show a clear voltage reduction, which becomes more pronounced as the discharge current increases. This behavior is particularly evident at medium and high current rates and is mainly associated with increased polarization, elevated internal resistance, and reduced Li-ion diffusion at low temperature. Under these conditions, the influence of the current rate on the voltage response is maximized, as higher currents amplify polarization effects throughout the SOC range.
At 25°C, the voltage-SOC characteristics become smoother and more stable, representing nominal operating conditions where electrochemical reactions proceed with well-balanced kinetics. The discharge current influence is still visible but is significantly reduced compared to low-temperature operating, indicating that the cell operates in its optimal electrochemical range, providing a reliable reference for performance and health assessment.
At 45°C, the discharge voltage increases across the SOC range and the spacing between curves obtained at different current rates becomes minimal. This behavior reflects a significant reduction in internal impedance, faster charge-transfer kinetics, and improved lithium-ion mobility within the electrode and electrolyte. As a results, the cell becomes significantly less sensitive to discharge current rate, even at high rate up to 2C, highlighting the dominant role of temperature in governing electrochemical performance under these conditions.
Overall, these three temperature conditions highlight how thermal effects strongly govern the voltage response of LFP cells, reinforcing the importance of including temperature as a key input in the predictive AI model.
The selected temperature (0°C,25°C, and 45°C) provide adequate variability for training the AI model, as they span the practical operating range of LFP cells. At low temperature (0°C) increased polarization and kinetic limitations strongly shape the voltage response, while 25°C represents nominal behavior. High temperature (45°C) reduces impedance and accelerates kinetics, generating distinct voltage-SOC patterns. This diversity is sufficient for the model to learn the key nonlinear interactions between temperature, current, and SOC.
Figure 3. Effect of current rate and temperature on the voltage-SOC characteristics of LFP cells
Figure 3 illustrates the voltage-SOC training curves obtained at different C-rate (0.05C, 0.5C, 1C and 2C) and temperatures (0°C, 25°C, and 45°C). At low current (0.05C), the voltage-SOC curves exhibit minimal polarization effects, and the influence of temperature is relatively limited, resulting in closely overlapping profiles across the entire SOC range. As the current increases, the impact of both current rate and temperature becomes more pronounced.
At moderate to high C-rate (0.5C to 2C), significant voltage depression is observed at low SOC, particularly at 0°C, reflecting increased internal resistance and kinetic limitations. Higher temperatures (25°C and 45°C) lead to higher voltage levels and smoother voltage-SOC transitions, indicating improved electrochemical kinetics and reduced polarization. Additionally, the divergence between temperature-dependent curves increases with current rate, highlighting the coupled effect of thermal and electrical stress on cell voltage behavior.
Overall, these results demonstrate that both current rate and temperature strongly influence the voltage-SOC characteristics of LFP cells. This confirms the necessity of incorporating multi-current and multi-temperature data in the training of the GPR-based reference model to ensure accurate and condition-aware voltage prediction across a wide operating envelope.
4. Gaussian Process Regression Model for Voltage-SOC prediction
A GPR model was developed to approximate the nonlinear mapping from operating conditions to terminal voltage, creating a probabilistic ‘digital twin’ of a new cell.
T represents temperature (°C), I the discharge current (Crate), SOC the state of charge (%), and V the discharge voltage (V). where
is Gaussian noise.
The function f(x) is assumed to be drawn from a GP prior defined by a mean function (often set to zero after data normalization) and a covariance kernel function k(x,x’).
The choice of kernel is crucial. A composite kernel such as a Radial Basis Function (RBF) is often effective. The RBF captures smooth global variations, while the Matern kernel can model local nonlinearities and roughness. Hyperparameters (length scales for each input, signal variance, a noise variance) re-optimized by maximizing the log marginal likelihood of the training data. This data_driven approach requires minimal prior knowledge of the battery’s internal electrochemistry.
he trained GPR model Mgpr predicts the expected voltage
and importantly, its associated uncertainty
for any iput tuple (SOC, I, T) within the trained domain. This uncertainty quantification is a key advantage of GPR.
The predictive capability of the trained GPR model is illustrated through voltage_SOC curves under different operating conditions. Figure 3 presents the comparison between the experimental measurements and the GPR-predicted curves for a new LFP cell at 25°C and 1C discharge current. The predicted curve closely follows the measured data, accurately capturing the initial voltage drop, the plateau region, and the final voltage decline. Quantitative analysis shows a high correlation (
and a low RMSE, confirming the model’s fidelity under nominal conditions.
Fig. 3
Voltage-SOC predicted vs. Experimental voltage at 45°C and 1C.
Figure 4 shows the model’s performance under more challenging conditions (0°C and 2C discharge current). Despite the pronounced nonlinear behavior, including voltage sag and plateau shift, the GPR-predict curve reproduces the key features of the experimental data. Deviations remain within acceptable limits, demonstrating the model’s robustness and its ability to generalize beyond the training dataset.
Fig. 4
Voltage-SOC predicted vs Experimental voltage
These results highlight the effectiveness of the GPR framework in accurately reproducing the voltage-SOC behavior of LFP cells across a wide range of operating conditions. The model provides a continuous reference curve, enabling subsequent diagnostic and SOH assessment without the need for additional experimental measurements.
5. Diagnostic system for cell health assessment
In this section, a diagnostic system is introduced to quantify capacity degradation in cycled LFP cells. The proposed method is based on experimental discharge-voltage measurements of degraded cells, collected at various operating conditions. Measured discharge data from aged cells, including current I and discharge time
are used to calculate the degraded cell capacity ‘
’:
2
Using the operating conditions extracted from the aged cell discharge tests namely the applied current and temperature, the GPR model reconstructs the expected voltage response of a new reference under identical conditions, establishing a reliable benchmark for assessing degradation. From the predicted voltage-time response, the end of discharge point is identified, and the corresponding discharge duration
is combined with the applied current I to compute the nominal capacity of the new cell ‘
’.
3
The capacity loss and SOH are then quantified by comparing the degraded and nominal capacities, providing a condition-normalized degradation metric defined as:
4
5
The workflow of the proposed diagnostic system, encompassing the acquisition of aged-cell discharge data, GPR_based reconstruction of the healthy voltage response, calculation of nominal and degraded capacities, and subsequent SOH estimation, is summarized in Figure 5.
Fig. 5
proposed diagnostic system
To support effective monitoring and rapid decision-making, a dedicated diagnostic interface has been developed. The interface (Fig. 6) offers a modular architecture for visualizing voltage-SOC and voltage-time relationships, facilitating rapid and accurate evaluation of battery health under varying temperature and current conditions. Its design supports seamless integration with the broader diagnostic system, enabling condition aware monitoring and data-driven decision-making.
Fig. 6
Diagnostic interface for cell health monitoring
6. Results and discussion
To validate the proposed diagnostic system, three cycled LFP cells were subjected to controlled discharge tests under the conditions summarized in Table 1. All cells were fully charged to 100% SOC prior to testing to ensure comparable initial states. Voltage data were recorded throughout the discharge process to generate detailed voltage-time profiles for each cell. These profiles were subsequently used to calculate the actual discharge capacity of each cycled cell, serving as a reference for diagnostic validation.
|
Cell
|
Temperature (C°)
|
Current (C_rate)
|
|---|---|---|
|
1
|
16
|
2C
|
|
2
|
32
|
1.2C
|
|
3
|
22
|
0.5C
|
Figure 7 illustrates the voltage-time profile of the three LFP cells, including both new and degraded states. These profiles provide a detailed temporal representation of the voltage response under the specific temperature and current conditions listed in Table 1. The measured data serve as a reference for evaluating the diagnostic system, enabling extraction of actual discharge capacity and assessment of the impact of aging and cycling on cell performance.
Fig. 7
Voltage-time profiles of new and degraded LFP cells under different temperature and current conditions
The experimentally determined capacities were then compared with the reference values predicted by the integrated GPR based diagnostic model. This comparison allowed for a quantitative evaluation of the model’s predictive accuracy under diverse temperature and current scenarios. Subsequently, the SOH and the capacity loss of each cell were computed, providing a comprehensive assessment of cell aging and degradation patterns. Specifically, cell1 exhibited a capacity loss of 4% with an SOH of 96%, cell 2 showed a 7% loss of capacity with an SOH of 93%, and cell 3 presented a 12% loss of capacity with an SOH of 88%. By combining both the measured data and the model predictions, the diagnostic system demonstrates its capability to reliably quantify the performance deterioration of cycled LFP cells across different operating conditions.
7. Conclusion
This study introduces an AI-based diagnostic methodology for aging characterization of LFP battery cells using GPR predictive reference model. By training the GPR model on experimental datasets collected from a new cell at 0°C, 25°C, and 45°C, we developed a predictive reference capable of estimating the voltage-SOC curve for any realistic operating condition.
The proposed diagnostic approach compares the predicted reference voltage of a new cell with the measured voltage of a cycled cell. The resulting deviation provides a direct and quantitative indication of degradation, enabling an accurate estimation of the cell’s SOH. This approach significantly provides a fast, efficient, and reliable tool for quality control and early detection of abnormal aging.
Experimental validation on three cycled LFP cells demonstrated the system’s ability to capture performance deterioration under diverse temperature and current conditions, with Cell1, Cell2 and Cell3 showing capacity losses of 4%, 7% and 12% (SOH = 96%,93%, and 88%, respectively). By integrating measured voltage-SOC and voltage-time profiles with GPR predictions, the framework offers a comprehensive, condition-aware assessment of cell aging without interrupting cycling or altering test conditions.
A modular diagnostic interface further supports visualization of voltage-SOC and voltage-time trends, facilitating rapid decision-making and in_situ monitoring. This approach significantly reduces testing time and operational complexity in laboratory and industrial settings, providing a scalable solution for applications in renewable energy storage systems, electric vehicles, and large-scale battery manufacturing.
Future work will focus on expanding the model to include additional degradation indicators, incorporating multi-cell variability and validating the methodology across different chemistries and usage profiles.
A
Author Contribution
Yahmadi Raja and Brik Kais: Conceptualization, methodology, data analysis, writing – original draft, and review & editing. Brik Kais: Validation and final review of the manuscript.
References
1.
P. Keil and A. Jossen, “Charging protocols for lithium-ion batteries and their impact on cycle life—An experimental study with different LFP cells,” Journal of Energy Storage, vol. 6, pp. 125–141, 2016.µ
2.
H. Gao et al., “Multi-factor aging in Lithium Iron Phosphate batteries: Mechanisms and insights,” Applied Energy, vol. 2024.
3.
Cai Qi, Tianshun Yao, Wenjie Zhai, Mingdao Zhang, Li Song, Jianping He,, “Advances in degradation mechanism and sustainable recycling of LiFePO4-type lithium-ion batteries,Energy Storage Materials,Volume 71,2024,103623, https://doi.org/10.1016/j.ensm.2024.103623.
4.
I. Azzouz, R. Yahmadi, K. Brik, F. Ben Ammar “Analysis of the critical failure modes and developing an aging assessment methodology for lithium iron phosphate batteries” Electrical Engineering, springer, ISSN 0948–7921, Volume 104, Issue 1, Pages 27–43February 2022, https://doi.org/10.1007/s00202-021-01320-7
5.
I. Azzouz, I. hammami, K. Brik, F. Ben Ammar “Integration of multi-criteria decision-making for performance evaluation of different solar batteries technologies” Electrical Engineering, springer, ISSN 0948–7921, Volume 105, Issue 2, Pages 775–795, April 2023, https://doi.org/10.1007/s00202-022-01697-z
6.
Junhui Cai, Yanjuan Li,* Shengnv Xu, Yiran Li, ZhanZhan Wang, Jie Liu, Shun Yang *andXiao Yan,”Failure mechanisms of LFP cathodes, Nanoscale, 2025 — étude sur les mécanismes de dégradation de cathodes LFP (perte de lithium, fissures, dissolution, etc.). Nanoscale, 2025, 17, 1204
7.
S. Saxena, C. Hendricks, and M. Pecht, “Cycle life testing and modeling of LFP cells under different aging stress factors,” Journal of Power Sources, vol. 327, pp. 395–406, 2016.
8.
H. Zheng et al., “A Review of Capacity Fade Mechanism and Promotion Strategies for Lithium Iron Phosphate Batteries,” Coatings, vol. 14, no. 7, p. 832, 2024.
9.
Y. Li et al., “Accelerated aging and degradation mechanism of LiFePO₄/graphite batteries cycled at high discharge rates,” RSC Advances, vol. 8, pp. 22456–22463, 2018.
10.
M. Berecibar et al., “Critical review of state-of-health estimation techniques of Li-ion batteries,” Renewable and Sustainable Energy Reviews, vol. 56, pp. 572–587, 2016.
11.
L. Cai and R. E. White, “Mathematical modeling of a lithium-ion battery with thermal effects in COMSOL Inc. Multiphysics (MP) software,” Journal of Power Sources, vol. 196, pp. 5985–5989, 2011.
A
12.J. Vetter et al., “Ageing mechanisms in lithium-ion batteries,” Journal of Power Sources, vol. 147, pp. 269–281, 2005.
Y. Che, X. Hu, X. Lin, J. Guo, R. Teodorescu, Health prognostics for lithium-ion batteries: mechanisms, methods, and prospects, Energy & Environmental Science, vol. 16, pp. 338–371, 2023. https://doi.org/10.1039/d2ee03019e
K. A. Severson, P. M. Attia, N. Jin, et al., Data-driven prediction of battery cycle life before capacity degradation, Nature Energy, vol. 4, pp. 383–391, 2019. https://doi.org/10.1038/s41560-019-0356-8
P. M. Attia, A. Grover, N. Jin, et al., Closed-loop optimization of fast-charging protocols using machine learning, Nature, vol. 578, pp. 397–402, 2020. https://doi.org/10.1038/s41586-020-1994-5
Y. Liu, B. Li, J. Li, et al., Artificial intelligence-driven real-world battery diagnostics, Energy and AI, vol. 18, 100419, 2024. https://doi.org/10.1016/j.egyai.2024.100419
X. Li, K. Song, M. Pecht, Can machine learning accelerate battery development?, Nature Reviews Materials, vol. 7, pp. 645–664, 2022. https://doi.org/10.1038/s41578-022-00473-7