Physics-Informed Machine Learning for Battery Degradation Diagnostics

Main codes for half-cell model, PINN and co-kriging implemented for physics-informed degradation diagnostics project: https://doi.org/10.1016/j.ensm.2024.103343

_{Note: The dataset used in this project is confidential and cannot be shared due to agreements with industry collaborators.}

PINN

A physics-informed neural network (PINN) model designed for battery degradation diagnostics. This model combines:

1️⃣ CustomLossHC Class

Defines the hybrid loss function, including:

• Data-driven MSE loss (between predictions and real data).
• Physics-based loss (constraining outputs using a half-cell model).
• Peak loss (minimizing dQ/dV peak differences using a surrogate model).

2️⃣ PINN Model

Implements a fully connected neural network (FCNN) with:

• Input layer → input_size neurons.
• Two hidden layers with ReLU activation.
• Output layer → output_size neurons.

3️⃣ PINNTrainer Class

Handles the training pipeline, including:

• Loading pre-trained surrogate and half-cell models.
• Training with Adam optimizer and early stopping.
• Batch-based training using PyTorch DataLoader.
• Validation to monitor performance.

Co-Kriging

Co-kriging is an extension of Gaussian process regression (GPR) that enables multi-fidelity modeling. This allows us to model a high-fidelity function using both high-fidelity (experimental) and low-fidelity (simulated by the half-cell model) datasets. This method utilizes a joint covariance function to simultaneously model the auto-covariances of each individual process and the cross-covariance between two related processes.

1️⃣ CoKrigingModel Class

The CoKrigingModel class handles training, optimization, and prediction.

Methods:

• __init__() → Initializes the model, loads data, and defines the kernels.
• get_data_and_split() → Loads the dataset and splits it into training/testing sets.
• train_model() → Optimizes the Co-Kriging model using maximum likelihood estimation.
• predict() → Generates predictions for new test inputs.
• run() → Trains the model, makes predictions, and calculates RMSPE.

2️⃣ test_cokriging() Function

This function serves as the main entry point to train and evaluate the co-kriging model.