Learning continuous state of charge dependent thermal decomposition kinetics for Li-ion cathodes using Kolmogorov-Arnold Chemical Reaction Neural Networks (KA-CRNNs)

This paper introduces a physics-encoded Kolmogorov-Arnold Chemical Reaction Neural Network (KA-CRNN) framework that learns continuous, interpretable state-of-charge-dependent thermal decomposition kinetics for Li-ion cathodes directly from DSC data, overcoming the limitations of existing models that rely on discrete or scalar kinetic parameters.

The Gist

The Big Picture: Predicting Battery Fires Before They Happen

Imagine your smartphone or electric car battery as a tiny, high-tech city. Inside this city, there are chemical "citizens" (atoms and molecules) constantly moving and interacting. Usually, they live in harmony. But if the city gets too hot or too charged up, these citizens can panic, start fighting, and cause a massive fire. This is called thermal runaway.

Scientists have long known that the "mood" of this city depends heavily on how full the battery is. If the battery is nearly empty (low charge), it's calm. If it's nearly full (high charge), it's on edge and much more likely to explode.

The Problem:
Until now, the "weather forecasts" (predictive models) scientists used to predict these fires were like checking the weather only once a day at noon. They could tell you what happens when the battery is 100% full, or maybe 50% full, but they couldn't smoothly explain what happens at 51%, 52%, or 53%. They treated the battery's behavior as a series of disconnected snapshots rather than a continuous movie. This made it hard to predict exactly when a battery might go critical during real-world driving or charging.

The Solution: A "Smart" Neural Network with a Memory

The authors of this paper, Benjamin Koenig and Sili Deng, created a new kind of AI tool called a KA-CRNN (Kolmogorov-Arnold Chemical Reaction Neural Network).

Think of this tool as a super-smart detective that doesn't just memorize facts; it understands the story behind the chemistry.

Here is how they built it, using a few creative analogies:

1. The "Recipe Book" (The Physics)

Usually, AI just guesses patterns. But this AI was given a "recipe book" based on real chemistry laws (like the Arrhenius equation). It knows that:

• Step 1: The battery's positive electrode (cathode) changes its shape as it gets hotter.
• Step 2: When it changes shape, it might release Oxygen (like a balloon popping).
• Step 3: That Oxygen rushes over to the liquid electrolyte (the fuel) and causes a fire.

The AI was told: "You must follow these steps. You cannot invent new physics." This ensures the predictions are scientifically real, not just math tricks.

2. The "Dial" (The State of Charge)

The big innovation is how this AI handles the State of Charge (SOC).

• Old AI: Had a fixed recipe. "If the battery is 100% full, use Recipe A. If it's 50%, use Recipe B." It couldn't handle the in-between.
• New AI (KA-CRNN): Has a continuous dial. Instead of switching recipes, it smoothly turns the knobs on the recipe. As you turn the dial from 50% to 51%, the AI subtly adjusts the "spiciness" (reaction speed) and "amount of oxygen" released. It creates a smooth, flowing movie of the battery's behavior rather than a slideshow.

3. The "Critical Tipping Point"

The researchers discovered something fascinating while training this AI. They found a Critical SOC.

• Analogy: Imagine a dam holding back water. As the water level rises, the dam holds fine. But at a specific height (say, 85% full), the pressure suddenly becomes too much, and the dam cracks instantly.
• The Discovery: The AI learned that for certain high-performance batteries (like Nickel-rich ones), there is a specific charge level where the battery suddenly stops being stable. Below this level, it releases a little oxygen. Above it, it releases a lot of oxygen, causing a rapid fire. The AI could pinpoint this exact "tipping point" for different battery types.

How They Tested It

They fed the AI data from a machine called a DSC (Differential Scanning Calorimeter), which basically heats up tiny battery samples and measures how much heat they give off.

They tested three different types of battery "cities" (NCA, NM, and NMA chemistries).

• The Result: The AI didn't just memorize the data; it learned the underlying rules. It could predict the heat release for a battery charge level it had never seen before (like predicting the weather for a Tuesday when it only trained on Mondays).
• The "Aha!" Moment: The AI correctly identified that for some batteries, the danger zone starts around 210–230 mAh/g (a specific measure of charge). It showed that the oxygen release jumps suddenly at this point, explaining why those batteries catch fire so easily when fully charged.

Why This Matters for You

This isn't just academic theory. This new model helps engineers:

1. Design Safer Batteries: By knowing exactly when the "tipping point" happens, engineers can design batteries that stay stable even when fully charged.
2. Better Safety Systems: Current safety systems might wait until a battery is 100% full to worry. This model says, "Hey, at 85% full, this specific battery type is already in the danger zone." This allows for smarter, real-time safety monitoring.
3. Interpretability: Unlike "black box" AI that gives an answer without explaining why, this model is transparent. We can look at the "knobs" it turned and say, "Ah, it increased the oxygen release rate because the battery was at 80% charge."

Summary

In short, the authors built a physics-aware AI that acts like a continuous movie camera for battery chemistry. Instead of taking blurry snapshots at specific charge levels, it captures the smooth, real-time story of how batteries heat up, release oxygen, and potentially catch fire. This allows us to predict and prevent battery fires with much higher precision, keeping our electric cars and phones safer.

Technical Summary

1. Problem Statement

Thermal runaway in lithium-ion batteries (LIBs) is heavily influenced by the State of Charge (SOC). Existing predictive models face significant limitations:

• Discrete vs. Continuous: Traditional models (e.g., Kissinger method) or existing Chemical Reaction Neural Networks (CRNNs) typically infer kinetic parameters as fixed scalars at specific, discrete SOC levels (often only 100% SOC). They fail to capture the continuous dependence of exothermic behavior on SOC.
• The "Critical SOC" Phenomenon: Experimental data (specifically Differential Scanning Calorimetry, DSC) reveals a critical SOC threshold (typically around 70–80% for nickel-rich cathodes) where thermal stability drops abruptly. Above this threshold, oxygen release and heat generation increase sharply. Current models cannot smoothly represent this transition or predict behavior at intermediate SOCs.
• Lack of Interpretability: While some recent approaches use neural networks, they often act as "black boxes" or rely on lumped reactants (e.g., Accelerating Rate Calorimetry data) that obscure specific component-level reaction pathways and mechanistic insights.

2. Methodology

The authors propose a Kolmogorov-Arnold Chemical Reaction Neural Network (KA-CRNN) framework to learn continuous, SOC-dependent kinetic parameters directly from DSC data.

A. Reaction Mechanism and Model Form

The study models the decomposition of nickel-rich cathodes (NCA, NM, NMA) mixed with electrolyte using a three-step global reaction pathway:

1. $R_1$ (Layered $\to$ Spinel): Phase transition of the cathode.
2. $R_2$ (Spinel $\to$ Rock Salt): Further decomposition releasing Oxygen ($O_2$).
3. $R_3$ (Electrolyte Oxidation): Reaction between released $O_2$ and the electrolyte.

The key innovation is that the kinetic parameters for $R_1$ and $R_2$ (activation energy $E_a$, pre-exponential factor $A$, reaction order $n$, enthalpy $\Delta H$, and crucially, the oxygen stoichiometric coefficient $\nu$) are treated as continuous functions of SOC, while $R_3$ parameters remain scalar (SOC-invariant) but depend on the $O_2$ produced by $R_2$.

B. KA-CRNN Architecture

• Kolmogorov-Arnold Networks (KANs): Instead of standard neural network weights, the model uses KANs where learnable parameters are activation functions.
• Chebyshev Polynomial Basis: The SOC-dependent parameters $p_i(SOC)$ are represented as a sum of Chebyshev polynomials:
  $$\phi_i(SOC) = \sum_{n=0}^{N} w_{i,n}^\psi \cdot \psi_n(SOC)$$
  where $\psi_n(SOC) = \cos(n \cdot \arccos(SOC))$.
• Hybrid Structure: The framework combines the mechanistic rigor of CRNNs (governed by Arrhenius laws and mass action) with the functional expressiveness of KANs. This allows the network to output real, interpretable kinetic parameters at any arbitrary SOC value.

C. Training and Optimization

• Datasets: Trained on DSC data for three cathode chemistries (NCA, NM, NMA) across nine discrete SOC levels (58.3% to 87.4%).
• Loss Function: An augmented Mean Squared Error (MSE) loss is used, incorporating physics-informed penalties:
  • Monotonicity ($L_{mono}$): Enforces that oxygen release ($\nu$) and reaction enthalpy ($\Delta H_2$) increase (or remain stable) as SOC increases, reflecting known physical degradation trends.
  • Stability Constraints ($L_{min}, L_{max}$): Prevents reaction orders and temperature factors from drifting to nonsensical values.
• Generalization: One dataset (72.9% SOC) was withheld for testing to verify the model's ability to interpolate and generalize without overfitting.

3. Key Contributions

1. Continuous Kinetic Modeling: First framework to learn continuous, fully interpretable kinetic parameters as functions of SOC, bridging the gap between discrete experimental points.
2. Mechanistic Interpretability: Unlike black-box models, the KA-CRNN explicitly learns the physical parameters (e.g., $E_a$, $\nu$) at every SOC. This allows researchers to visualize exactly how the oxygen release stoichiometry ($\nu$) changes, revealing the "critical SOC" transition point.
3. Unified Multi-Chemistry Model: Successfully trained a single framework across three different cathode chemistries (NCA, NM, NMA), capturing their unique stability profiles while sharing a universal electrolyte oxidation reaction ($R_3$).
4. Physics-Guided Learning: By embedding the reaction pathway and physical constraints (monotonicity) directly into the loss function, the model avoids spurious oscillations and ensures physical consistency even with sparse data.

4. Results

• DSC Reconstruction: The model accurately reproduced DSC heat-release curves across all SOCs for all three cathodes, including the withheld testing data. It successfully captured the shift in peak temperature, peak height, and total heat release as SOC increased.
• Critical SOC Identification: The learned oxygen stoichiometric coefficient ($\nu$) showed a sharp, continuous increase at specific SOC thresholds (e.g., ~210–230 mAh/g), directly correlating with the sudden spike in heat release observed in experiments.
• Chemical Insights:
  • NM Cathode: Showed the lowest thermal stability, with early onset of oxygen release and high heat generation.
  • NMA Cathode: Demonstrated higher stability (higher critical SOC) and a broader $R_2$ peak, which the model correctly interpreted as evolving oxygen over a wider temperature window, thereby limiting the rate of the subsequent electrolyte oxidation ($R_3$).
  • Parameter Trends: The model revealed that while $R_1$ parameters were relatively stable, $R_2$ parameters (activation energy and pre-exponential factor) shifted significantly near the critical SOC, driving the instability.

5. Significance and Future Outlook

• Safety Monitoring: This approach enables real-time thermal runaway prediction and monitoring that adapts to the battery's current SOC, moving beyond conservative "worst-case" (100% SOC) models.
• Scalability: The modular nature of KA-CRNNs allows for the extension of this framework to other variables (e.g., pressure, temperature, aging) and other battery components (anodes, SEI layers).
• Uncertainty Quantification: The authors suggest future integration with Bayesian CRNNs to handle experimental uncertainty and provide probabilistic safety assessments.
• Data Efficiency: The framework demonstrates that high-fidelity, physics-based models can be learned from relatively sparse experimental data (9 SOC points) by leveraging strong physical priors and continuous function approximation.

In summary, this work establishes a new paradigm for battery safety modeling, replacing static, discrete kinetic parameters with dynamic, continuous, and physically interpretable functions that accurately reflect the complex, SOC-dependent nature of thermal runaway.