Session: 09-01-01: Electrochemical Energy Storage and Conversion Systems I

Paper Number: 173480

Data-Driven Safety Analytics for Lithium-Ion Batteries 

Accurate battery thermal safety predictions are critical for optimizing the reliability and efficiency of energy storage systems, particularly under varying operating conditions. The complex interplay of electrochemo-mechanical-thermal degradation mechanisms often limits the predictive accuracy of both physics-based and machine learning models. In this study, we demonstrate a comprehensive framework for evaluating both electrochemical and thermal stability of lithium-ion batteries. This study employs in-house samples of lithium-ion cells. These samples are developed for thermal stability assessment, investigated using accelerating rate calorimetry (ARC) to predict key thermal safety parameters during thermal runaway. Leveraging an eclectic mix of machine learning techniques, we interrogate key factors of prediction accuracy and demonstrate an adaptable prognostic framework capable of forecasting the thermal safety of lithium-ion cells across a wide range of cycle counts and form factors. Our framework approach underscores the potential of physics-informed data-driven models to enhance battery prognostics, providing a scalable and efficient pathway for optimizing energy storage reliability and safety.

Our study employs accelerating rate calorimetry (ARC) experiments on CR2032 coin cells containing nickel-cobalt-manganese-aluminum (NCMA) cathodes and silicon oxide-doped graphite anodes to assess their thermal safety behavior. These cells exhibit differing state-of-charge (SOC) levels and electrode compositions to reflect real-world heterogeneity. From the ARC data, we extract five critical thermal safety metrics: self-heating onset temperature, thermal runaway onset temperature peak runaway temperature , maximum self-heating rate ,  and total adiabatic heat release (Q). These metrics provide a holistic view of the cell’s thermal behavior during abuse scenarios. Crucially, the model demonstrates robust predictive capabilities across diverse conditions, including unknown states-of-charge (SOC) and unknown anode compositions (silicon vs. graphite). Input features include metadata and specific experimental ARC observations, enabling reliable thermal safety predictions despite the limited availability of training data. A key component of our framework is the Random Forest Regression model, enhanced with second-order polynomial feature expansion, which enables it to capture complex nonlinear interactions among metadata and ARC-derived observations. The proposed approach is evaluated against several widely used machine learning baselines to ensure robustness and generalizability. Across both zero-shot (unseen SOC or anode composition) learning setups, the Random Forest model consistently outperforms its counterparts.

The study employs data from ARC experiments to analyze and predict critical thermal safety characteristics such as thermal runaway onset temperatures and exothermic heat release while demonstrating the model’s ability to generalize across varying SOC and anode compositions. These findings validate the utility of ensemble learning with polynomial augmentation as a powerful and interpretable strategy for safety-critical battery modeling. Our ML-based pipeline provides a scalable and accurate alternative to traditional thermal models, paving the way for advanced diagnostic tools and design optimization strategies. This work highlights the transformative role of data-driven methods in accelerating the development of inherently safer batteries and informing industrial practices and regulatory frameworks as the world transitions toward a more electrified and energy-resilient future.

 

Presenting Author: Pretty Mitra Purdue University

Presenting Author Biography: Pretty Mitra is a third-year PhD student in Mechanical Engineering at Purdue University. She is a Graduate Research Assistant in the Energy and Transport Sciences Laboratory (ETSL), where her work focuses on thermal safety analytics and data-driven modeling of batteries.

Authors:

Pretty Mitra Purdue University
Avijit Karmakar Electrochemical Safety Research Institute (ESRI), Underwriters Laboratories Inc.
Bairav S. Vishnugopi Purdue University
Partha P. Mukherjee Purdue University