Physics-Informed Machine Learning for Pouch Cell Temperature Estimation

1. Problem Statement

The rapid transition to transportation electrification has increased the reliance on high-performance pouch cells, which require precise thermal management to ensure safety, efficiency, and longevity.

• The Challenge: Accurately estimating temperature distributions within pouch cells equipped with indirect liquid cooling (cold plates) is critical for design optimization.
• Limitations of Current Methods:
  • Finite Element (FE) Analysis: While accurate, FE simulations are computationally expensive, making them impractical for iterative design optimization and real-time applications.
  • Purely Data-Driven Models: Traditional deep learning approaches rely heavily on large, diverse datasets. They often struggle to generalize when data is sparse and may violate fundamental physical laws (e.g., energy conservation), leading to implausible predictions, especially in regions far from cooling channels.

2. Methodology

The paper proposes a Physics-Informed Machine Learning (PIML) framework that integrates governing heat transfer equations directly into the neural network's training process.

A. Problem Formulation & Data Generation

• System: A 2D steady-state heat conduction problem representing a pouch cell ($154 \times 203$ mm) with a thin cold plate.
• Governing Equation: The model is based on the steady-state 2D heat diffusion equation with internal heat generation and localized convective cooling:
  $$k \cdot t \cdot \nabla^2 T + q_{gen} - h(x, y) \cdot (T - T_{coolant}) = 0$$
  Where $h(x,y)$ varies based on a binary channel mask (high convection under channels, zero elsewhere).
• Ground Truth: A dataset of 100 unique cooling channel configurations was generated using the Finite Difference Method (FDM) to serve as the "ground truth" for training and validation.

B. Comparative Approaches

The study compares two modeling strategies:

1. Data-Driven Surrogate Model:
   • A Fully Convolutional Network (FCN) trained on input channel masks ($X$) to predict temperature fields ($T$).
   • Loss function: Mean Squared Error (MSE) between predicted and normalized ground truth.
   • Training: 100 epochs using the Adam optimizer.
2. Physics-Informed Machine Learning (PIML) Model:
   • Architecture: A Physics-Informed Neural Network (PINN) that takes continuous spatial coordinates $(x, y)$ as input and outputs temperature $T(x, y)$.
   • Loss Function: A composite loss function ($L_{total}$) combining three terms:
     $$L_{total} = w_1 \cdot L_{MSE} + w_2 \cdot L_{PDE} + w_3 \cdot L_{BC}$$
     • $L_{MSE}$: Data fidelity (difference from FDM ground truth).
     • $L_{PDE}$: Residual of the governing partial differential equation (ensuring the solution satisfies heat transfer physics).
     • $L_{BC}$: Boundary condition penalty (enforcing adiabatic conditions at outer edges).

3. Key Contributions

• Framework Integration: Successfully embedded the specific 2D heat transfer PDE and boundary conditions of a pouch cell cooling system directly into the neural network loss function.
• Efficiency: Demonstrated that PIML can achieve high-fidelity predictions with significantly fewer training samples and faster convergence compared to purely data-driven methods.
• Generalization: Showed that incorporating physical constraints allows the model to perform robustly in regions where data is sparse (e.g., areas away from cooling channels), overcoming a major weakness of standard deep learning.

4. Results

The models were evaluated on a dataset of varying cooling channel geometries:

• Convergence Speed: The PIML model converged much faster. After just 10 epochs, the PIML model achieved a Mean Squared Error (MSE) of 5.66, whereas the data-driven model had an MSE of 11.12.
• Accuracy Improvement: The PIML approach resulted in a 49.1% reduction in MSE compared to the purely data-driven model.
• Spatial Accuracy: Visual analysis (Fig. 2) confirmed that the PIML model provided more accurate temperature estimations, particularly in regions outside the immediate path of the cooling channels, where purely data-driven models tended to fail or produce less reliable gradients.

5. Significance and Future Work

• Surrogate Modeling: This work establishes PIML as a superior surrogate modeling tool for Battery Thermal Management Systems (BTMS). It offers a pathway to reduce reliance on expensive high-fidelity simulations and extensive experimental testing.
• Design Optimization: The speed and accuracy of the PIML framework enable rapid iterative optimization of cooling channel designs, which is crucial for developing next-generation EV batteries.
• Future Directions: The author plans to extend this framework to optimize advanced cooling channel geometries and validate the model against physical experimental data to further confirm its predictive capabilities in real-world scenarios.

In summary, this paper demonstrates that integrating physical laws into machine learning is not just a theoretical improvement but a practical necessity for achieving reliable, fast, and accurate thermal management in electric vehicle battery systems.