Physics-Informed machine learning for Li-ion battery State of Health forecasting via degradation modes

The prediction of the State of Health (SoH) of lithium-ion batteries has been extensively studied, but rarely under the scope of the underlying Degradation Modes (DM). Yet, the identification of DM is essential to achieve better diagnostics and informed battery management strategies. To address this challenge, this study proposes a novel machine learning (ML) framework for SoH forecasting, leveraging DM as intermediate variables. The proposed architecture is composed of two distinct ML models, the first one is dedicated to DM forecasting, while the second maps these predictions to SoH estimates. The approach constrains the neural networks within a physically meaningful latent space, enhancing interpretability and performance. To train the initial model, three heterogeneous categories of input features were engineered. The first one consists of latent representations derived from the Incremental Capacity Analysis (ICA) curves, using a Variational AutoEncoder (VAE), the second one comprises high-level features extracted from charging curves using statistical methods, and the third one captures the historical evolution of degradation. The combination of these inputs provides complementary information, enhancing the model's forecasting capability. Prior knowledge of DM evolution is used to constrain the loss function, guiding the learning process towards physically consistent predictions. The proposed architecture demonstrates the capability to forecast up to 14% of the total cycle lifetime utilizing 16% of historical cycle data, while maintaining a SoH prediction error below 5%.

Keywords: Degradation Modes, Machine Learning, Li-ion battery, State of Health, Deep Learning, physics-informed