CAREER: Active Learning of Second-Life Battery Systems by Combining Reinforcement Learning Principle and Device Physics

This Faculty Early Career Development (CAREER) project aims at developing a transformative active learning framework, which is critical for enabling second-life battery and other energy system applications. Reuse/repurposing of retired electric vehicle (EV) batteries has been considered as a critical approach for facilitating transportation electrification and renewable power generation, which are two cornerstones of the emerging clean energy revolution essential to our national prosperity.

EV batteries are mandated to retire with 80% remaining capacity, and direct disposal would mean a substantial waste of the remaining value. Repurposing these batteries for the less-demanding stationary storage, e.g. to store the renewable but intermittent solar/wind energy, could significantly improve the cost and sustainability of both EV and renewable power generation industries.

A key challenge facing repurposing is the risk of unsafe/unhealthy battery operation due to the damage and degradation suffered from the first use. Therefore, current repurposing practice requires tedious and costly manual testing and grading of each retired battery module. Still, such one-time testing can only guarantee safety and performance at the beginning but not during subsequent operation.

A key enabler for battery repurposing at a much larger scale is the advanced technique for accurate, fast, automatic, and continuous estimation of battery states and parameters. The research component of this project will be integrated with an education plan with the theme “Connecting Emerging ML/AI with Traditional Control” to achieve the PI’s overarching education goal of promoting machine learning (ML) and artificial intelligence (AI) education among engineering students and professionals, especially underrepresented minorities.

The research will explore active learning of second-life batteries, where the input current is regulated to optimize the information content of the response battery voltage to improve the speed and accuracy of estimation. The key innovation is a generic active learning framework, which combines the principle of reinforcement learning with device physics to overcome the fundamental limitations in the current practice of active learning.

This is among the first attempts to use reinforcement learning for estimation, which presents a series of fundamental research problems in information state computation, reward and learning architecture design, and convergence analysis. The second part of the research focuses on promoting the learning performance for multi-modular systems through module collaboration.

The goal is to address fundamental challenges of reinforcement learning, e.g. the need for massive training data and slow convergence, by enabling data sharing and cooperative search among modules.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.