Advanced ScienceSource publication:
Loading content…
An LLM is used as a representation generator that emits executable battery fast-charging control policies, improving simulated state of health after 100 cycles by about 4.2% over a multi-step constant-current baseline.
The design space for battery fast-charging protocols no longer has to be fixed in advance by a researcher, because a language model generates executable control-policy code. Earlier charging-protocol optimization usually restricted the search to preset parameterizations such as exponential or sigmoid curves or a small number of piecewise-constant segments, which are easy to optimize but limit the diversity of strategies that can be found. In simulations using a DFN electrochemical model coupled to thermal and degradation submodels, both language-model methods beat a three-segment constant-current baseline implemented after Attia et al., raising final state of health by 4.24% and 4.17%.
The language model does not perform numerical training itself; it proposes either the structure of a policy or a complete expression, and the battery evaluator decides how good it is. Language models are unreliable at arithmetic and quantitative precision, so numerical parameters are either handed to the gradient-free SAASBO optimizer or treated only as candidates that the evaluator screens. Prompt-to-Optimizer has an outer language model generate small networks with fewer than 35 trainable parameters and an inner SAASBO loop tune them, while Prompt-to-Protocol has the language model emit a fully specified protocol; the two are compared under a matched budget of 210 full protocol evaluations.
The results come from PyBaMM simulations with a DFN electrochemical model plus a lumped thermal model and degradation submodels, for a 2.5 Ah cell charged from 0 to 90% SOC in 0.5 h at roughly 1.8 C on average, with all methods compared under the same budget of full protocol evaluations. The intended reader is someone working on battery protocol optimization or gradient-free black-box optimization; applying a protocol to a physical cell would still require model calibration, safety screening, and cell-level validation.
The language model is explicitly not treated as a precise numerical optimizer, and the success of Prompt-to-Protocol is interpreted as black-box candidate generation rather than precise numerical training. The generated expressions are problem-specific heuristics that may not transfer directly to different cells or operating conditions; the simulations do not isolate the contribution of individual degradation mechanisms and do not include all harmful side reactions, so an additional voltage penalty stands in for unmodeled high-voltage processes.