Reproduce “PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction” on OpenResearch

About this paper

Zeping Chen, Ruda Jian, Sachin Sigdel, Guoping Xiong, Jian-Xun Wang, Tengfei Luo

Researchers at the University of Notre Dame and The University of Texas at Dallas developed PiDDM, a physics-informed differentiable degradation modeling framework for lithium-ion battery State-of-Health (SOH) prediction. This approach integrates empirical degradation kinetics into a neural network, yielding superior prediction accuracy with an average Mean Squared Error of 6.6271e-5 in standard prediction and 1.1067e-4 in temporal extrapolation, while ensuring physically consistent and monotonic capacity fade.

Key insights

• Integrating empirical degradation kinetics as a physics-informed inductive bias significantly enhances predictive accuracy and generalization capabilities across diverse and dynamic operating conditions.
• Enforcing physical consistency through monotonic capacity fade within the degradation model effectively prevents physically implausible predictions, such as capacity regeneration, often observed in unconstrained data-driven models.
• The differentiable degradation module bridges mechanistic understanding with data-driven learning, allowing for robust extrapolation and accurate capture of complex, non-linear degradation behaviors like the end-of-life knee.

Summary by alphaXiv.