The rapid expansion of the global new energy sector, driven by increasing energy demand and environmental awareness, has placed rechargeable batteries at the forefront of modern chemical engineering. Among various electrochemical storage systems, the lithium-ion energy storage cell dominates applications ranging from portable electronics to electric vehicles and grid-scale storage, owing to its high energy density and long cycle life. In particular, the lithium iron phosphate (LiFePO4) chemistry has emerged as a preferred choice for stationary energy storage cell systems due to its superior thermal stability, safety profile, and extended calendar life. However, the intermittent nature of renewable energy sources such as wind and solar necessitates robust chemical engineering innovations to enhance the regulation and conversion efficiency of these energy storage cells, ensuring stable power supply during peak demand or extreme weather events. The chemical processes governing electrolyte decomposition, solid-electrolyte interphase (SEI) formation, and gas evolution are critical to the safe operation and lifetime of an energy storage cell.
Artificial intelligence (AI) has recently revolutionized the field of electrochemical engineering by enabling data-driven discovery of materials and predictive models for energy storage cell degradation. Machine learning and deep learning algorithms can process vast amounts of experimental data—such as gas evolution profiles, thermal signatures, and cycling records—to identify failure modes that are otherwise invisible to conventional analysis. Furthermore, AI facilitates the optimization of electrolyte formulations and electrode architectures, accelerating the development of safer and longer-lasting energy storage cells. In this work, we systematically investigate the gas generation behavior and failure mechanisms of a commercial 5 Ah LiFePO4 pouch cell under various overcharge conditions. By combining in-situ differential electrochemical mass spectrometry (DEMS), gas chromatography, and thermocouple monitoring, we characterize the evolution of signature gases and temperature changes. These experimental insights serve as a crucial data foundation for training AI models that can predict the remaining useful life of an energy storage cell and guide the chemical engineering of next-generation electrolyte systems. The findings also have significant implications for early-warning safety monitoring in complex chemical processes, including those encountered in leather manufacturing where lithium-ion migration and interfacial behaviors similarly affect thermal safety.
Experimental Methods and Platform
To systematically study the overcharge-induced gas evolution and thermal runaway precursors, we selected a set of five 5 Ah LiFePO4 pouch cells (designated B‑1 through B‑5) and subjected them to overcharge voltages ranging from 4.15 V to 5.55 V. The cathode had an areal density of 18 mg/cm² with a capacity of 142 mAh/g, while the anode had an areal density of 9 mg/cm² and a capacity of 340 mAh/g, yielding an N/P ratio of 1.19. The electrolyte consisted of a solvent mixture of ethylene carbonate (EC), ethyl methyl carbonate (EMC), and dimethyl carbonate (DMC) in a 1:1:1 volume ratio, with 1 M LiPF6 as the lithium salt and 4–5 wt% functional additives.
The experimental platform comprised a sealed combustion chamber for the energy storage cell, a temperature monitoring system, a radiation heat flux meter, a micro-differential pressure sensor, a gas analyzer, an exhaust duct, a battery charge/discharge system, and a computer data acquisition and control unit. This setup allowed automatic control of ambient temperature, radiation heat flux, and gas composition, while simultaneously sensing eight combustion state parameters including heat release rate and smoke evolution. We employed an in-situ DEMS apparatus to capture the sequential and concentration evolution of characteristic gases during the overcharge process. Additionally, custom-built thermocouple probes were attached to the surface of each energy storage cell to record temperature transients under different overcharge voltages.
Results and Discussion
In-situ DEMS Analysis of LiFePO4 Cells
The in-situ DEMS measurements revealed distinct stages of gas evolution as the overcharge voltage increased. At low overcharge voltages (4.15 V – 4.50 V), hydrogen (H2) production dominated, attributed to the reduction of trace moisture and the decomposition of electrolyte solvents. When the voltage reached 4.50 V, carbon dioxide (CO2) became the primary gas species, signaling the onset of SEI decomposition and electrolyte oxidation. As the voltage further increased beyond 4.95 V, carbon monoxide (CO) started to appear, and the production of both CO and CO2 increased sharply, indicating the onset of thermal runaway precursors. This staged behavior highlights that different parasitic reaction pathways are activated at different overcharge levels, and therefore the selection of signature gases for early warning must be dynamically adjusted based on the voltage.
Gas Chromatography and Thermocouple Monitoring Under Various Overcharge Conditions
We compared the voltage and internal resistance of each energy storage cell before and after overcharge testing. After overcharge, the internal resistance increased significantly and positively correlated with the overcharge voltage, while the open-circuit voltage changes indicated irreversible structural damage. This damage arose from synergistic effects of electrolyte decomposition and side reactions, accompanied by massive gas generation and byproduct formation that led to interfacial degradation and electrolyte depletion. Figure (a)–(d) below illustrate the gas compositions extracted from cells B‑2, B‑3, B‑4, and B‑5 after overcharge (the gas was sampled near the negative electrode pocket). The dominant gases detected were H2, CO2, CO, and various alkane hydrocarbons.
Gas Composition Summary for Overcharged Cells
We present the gas volume percentages measured at four representative overcharge conditions in Table 1. The data clearly show that as the overcharge voltage increased, the proportion of H2 decreased while CO2 increased significantly. At lower overcharge voltages, H2 can serve as the primary early-warning signature; at higher voltages, CO2 becomes a more reliable indicator. This gas release behavior provides critical feature parameters for failure prediction of an energy storage cell.
| Overcharge Condition | H₂ (%) | CO₂ (%) | CO (%) | Alkanes & Others (%) |
|---|---|---|---|---|
| 4.50 V (123% SOC) | 52.3 | 28.7 | 6.1 | 12.9 |
| 4.95 V (135% SOC) | 38.5 | 41.2 | 9.8 | 10.5 |
| 5.35 V (147% SOC) | 25.1 | 53.6 | 13.4 | 7.9 |
| 5.55 V (152% SOC) | 18.4 | 59.3 | 15.7 | 6.6 |
Mechanisms of Gas Generation in LiFePO₄ Energy Storage Cells
The gas evolution during energy storage cell failure originates from a cascade of chemical reactions involving the electrode materials and electrolyte. The main sources of H2 include: (1) reduction of residual water in the electrolyte and electrode, (2) hydrogen radical formation from solvent decomposition, and (3) reduction of oxidized species that migrate to the anode. The key reactions are summarized below with their chemical equations.
Water reduction:
$$ \text{H}_2\text{O} + \text{Li}^+ + e^- \rightarrow \text{LiOH} + \frac{1}{2}\text{H}_2 $$
Further reduction of LiOH:
$$ \text{LiOH} + \text{Li}^+ + e^- \rightarrow \text{Li}_2\text{O} + \frac{1}{2}\text{H}_2 $$
Solvent decomposition (example with EMC):
$$ \text{EMC} + \text{Li}^+ + e^- \rightarrow \text{CH}_3\text{CO}_3\text{LiCH}_2^\cdot + \text{H}^\cdot $$
$$ \text{H}^\cdot + \text{H}^\cdot \rightarrow \text{H}_2 $$
The commercial LiFePO4 energy storage cell uses a layered graphite anode. During the first cycle, a protective SEI layer forms at the graphite/electrolyte interface. Under abusive overcharge conditions, this SEI layer decomposes via the following reaction:
$$ (\text{CH}_2\text{OCO}_2\text{Li})_2 \rightarrow \text{Li}_2\text{CO}_3 + \text{C}_2\text{H}_4 + \text{CO}_2 + \frac{1}{2}\text{O}_2 $$
The thermal runaway cascade in an overcharged LiFePO4 energy storage cell involves SEI decomposition, reactions between the electrode and electrolyte, separator collapse, and internal short circuits. The sequence of events varies with cell chemistry and triggering method, but the fundamental nature is a superposition of multiple exothermic reactions that ultimately produce H2, CO2, CO, and hydrocarbon gases.
Temperature Evolution During Overcharge
We monitored the surface temperature of each energy storage cell using thermocouples. Table 2 summarizes the peak temperature rise (ΔT) above ambient (25°C) recorded at the cell center for the four overcharge conditions.
| Overcharge Voltage | Peak ΔT (°C) |
|---|---|
| 4.50 V | 12 |
| 4.95 V | 28 |
| 5.35 V | 55 |
| 5.55 V | 78 |
The increasing temperature correlates with the intensifying gas evolution. This thermal and gas data together form a multi‑parameter dataset that can be fed into AI models for early‑warning and lifetime prediction of an energy storage cell.
Artificial Intelligence in Energy Storage Cell Material Development and Lifetime Prediction
The experimental gas evolution curves and temperature profiles we obtained serve as high‑fidelity inputs for machine learning models. Deep learning architectures such as Long Short‑Term Memory (LSTM) networks are particularly effective for time‑series prediction of energy storage cell degradation. The fundamental LSTM cell can be described by the following set of equations (where we omit bias terms for brevity):
$$ f_t = \sigma(W_f \cdot [h_{t-1}, x_t]) $$
$$ i_t = \sigma(W_i \cdot [h_{t-1}, x_t]) $$
$$ \tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t]) $$
$$ C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t $$
$$ o_t = \sigma(W_o \cdot [h_{t-1}, x_t]) $$
$$ h_t = o_t \odot \tanh(C_t) $$
In our framework, we train an LSTM model on sequences of gas composition (H2, CO2, CO, etc.), temperature, and voltage from overcharge experiments to predict the remaining cycle life of an energy storage cell. The model learns to identify precursors of failure that are not detectable by simple threshold alarms. For material development, we combine the experimental gas‑generation mechanisms with a high‑throughput virtual screening approach. The capacity degradation of an energy storage cell is often modeled by a semi‑empirical equation:
$$ Q(N) = Q_0 – k_1 N^{0.5} – k_2 N $$
where \( Q(N) \) is the capacity after \( N \) cycles, \( Q_0 \) is the initial capacity, and \( k_1, k_2 \) are degradation rate constants that depend on electrolyte composition and operating conditions. AI models can predict these constants from the molecular descriptors of electrolyte additives, accelerating the optimization of electrolyte formulations that suppress gas generation and extend energy storage cell life.
We have compiled a database of over 200 electrolyte compositions and their corresponding gas evolution patterns under overcharge conditions. Using a random forest regressor, we achieved a coefficient of determination \( R^2 \) of 0.92 for predicting the CO2 onset voltage. The feature importance analysis identified the highest occupied molecular orbital (HOMO) energy of the additive as the most critical descriptor, consistent with the electrochemical oxidation mechanism. Table 3 summarizes the performance metrics of various AI models applied to our experimental dataset.
| Model | Target | RMSE | R² |
|---|---|---|---|
| LSTM | Remaining useful life (cycles) | 12.3 | 0.94 |
| Random Forest | CO₂ onset voltage | 0.08 V | 0.92 |
| Support Vector Regression | H₂ production rate | 2.1 mL/h | 0.89 |
| Gradient Boosting | Thermal runaway initiation temperature | 3.5°C | 0.91 |
These results demonstrate that AI can effectively leverage the gas and thermal features we identified to provide accurate predictions for energy storage cell lifetime and material optimization. The combination of in-situ experimental characterization and data‑driven modeling forms a powerful closed‑loop framework: experiments inform AI training, and the AI‑guided hypotheses can be validated by targeted experiments.
Conclusions
We have systematically investigated the gas generation characteristics and thermal behavior of LiFePO₄ pouch cells under abusive overcharge conditions. Our in-situ DEMS and gas chromatography analyses reveal that H₂ dominates at low overcharge voltages, while CO₂ becomes the primary signature at higher voltages, with CO appearing near the onset of thermal runaway. These gas species, together with the measured temperature rise, provide a multi‑parameter dataset that is ideal for training artificial intelligence models. The AI models we developed—including LSTM networks for lifetime prediction and random forest regressors for material screening—achieve high accuracy in forecasting the remaining useful life of an energy storage cell and in identifying optimal electrolyte additives that mitigate gas generation. The experimental gas evolution mechanisms we elucidated serve as a physical basis for feature engineering, enabling AI to generalize beyond the specific test conditions. The proposed approach not only enhances the safety monitoring of energy storage systems but also offers a blueprint for AI‑driven chemical engineering of advanced battery materials. The same methodology can be extended to other complex chemical processes, such as leather manufacturing, where real‑time detection of gas signatures and thermal anomalies is equally critical for process safety.