Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation

Researchers introduce Battery-Sim-Agent, an LLM-based framework that uses AI agents to estimate battery parameters by mimicking scientific reasoning rather than traditional black-box optimization. The system outperforms conventional methods like Bayesian optimization on benchmark tests and demonstrates practical applicability on real-world battery datasets, representing a novel approach to accelerating battery innovation through physics-informed AI reasoning.

Battery parameter estimation represents a fundamental bottleneck in battery development, requiring extensive computational resources and domain expertise to create accurate digital models. Traditional approaches rely on black-box optimization algorithms that treat the problem as a mathematical puzzle without leveraging physical intuition, resulting in sample inefficiency and extended development cycles. Battery-Sim-Agent reimagines this challenge by deploying large language models as reasoning agents that interpret simulator feedback, formulate physically grounded hypotheses, and iteratively refine parameters—essentially automating the trial-and-error process a human scientist would employ.

This work sits at the intersection of scientific computing and AI, where LLMs are increasingly being applied to domain-specific problems requiring structured reasoning. The framework's ability to handle multi-modal feedback and long-horizon degradation fitting suggests LLM agents can capture complex physical relationships that traditional optimization struggles with. By systematically outperforming Bayesian optimization baselines across diverse battery chemistries and operating conditions, the research demonstrates that reasoning-based approaches may be fundamentally more efficient for inverse problems in materials science.

The implications extend beyond batteries to broader scientific discovery workflows. If LLM-agents can reliably accelerate parameter estimation for complex physical systems, this could compress development timelines across battery, thermal, and electrochemical research. For the energy storage industry specifically, faster parameter identification could reduce costs and accelerate commercialization cycles. The validation on real-world datasets rather than purely synthetic benchmarks strengthens the practical relevance. Moving forward, the key question is whether this approach scales to higher-dimensional parameter spaces and whether the method generalizes across different battery simulator platforms and chemistries beyond the tested benchmarks.

• →LLM-agents outperform traditional black-box optimization methods like Bayesian optimization for battery parameter estimation
• →The framework successfully interprets multi-modal simulator feedback to form physics-grounded hypotheses and propose parameter updates
• →Battery-Sim-Agent demonstrates practical applicability on real-world battery datasets beyond synthetic benchmarks
• →This approach could significantly accelerate battery innovation cycles by automating human scientist workflows
• →The methodology potentially generalizes to other scientific discovery and inverse problem domains requiring parameter estimation