A language model writes battery fast-charging policies that, in simulation, keep about 4 points more battery health than a multi-step protocol

Ge Lei, Ferran Brosa Planella, Sterling G. Baird and Samuel J. Cooper published the paper in Advanced Science on 11 October 2026, open access under CC BY 4.0; a preprint went up on arXiv on 14 January 2026 as “From Prompt to Protocol”. Lei and Cooper are at Imperial's Dyson School of Design Engineering, Brosa Planella at the Faraday Institution and the University of Warwick, Baird at the University of Toronto's Acceleration Consortium.^[S1][S2]

The paper proposes two routes. P2P (Prompt-to-Protocol, one loop) has the language model write a complete charging policy as code; each candidate is simulated and scored, and the good ones stay in a pool as parents for the next round. P2O (Prompt-to-Optimizer, two loops) has the model write a small neural network as the policy's form; an inner loop tunes its parameters with SAASBO, a Bayesian optimizer that needs no gradients, and an outer loop has the model revise the architecture from the scores. Every experiment used GPT-5 (the 2025-08-07 version, medium reasoning effort); the prompts give only operating limits such as voltage, current and time, not the cell's chemistry or parameters. The authors say no generated code was corrected by hand and over 99% of candidates ran as written.^[S1]

Three cases run from simple to realistic: first, matching a predefined heat profile in a simple equivalent-circuit model, to test only the outer architecture search; second, finding the current that holds heat generation at a constant 0.4 W through a one-hour charge in a DFN electrochemical model, to test gradient-free tuning; third, the real fast-charging problem, in a DFN model with thermal and ageing submodels (SEI growth, particle cracking and others), charging from 0 to 90% in 30 minutes and reading battery health after 100 cycles.^[S1]

92.17% / 92.10% / 87.93%

final state of health of the best P2O, P2P and three-step baseline results in case three^[S1]

Limits in words make candidates behave from the start

The authors put P2P's edge at a matched budget down to the prompt: it states that going above 4.2 V is penalised, so candidates tend to respect the voltage limit before any numerical evaluation. P2O's neural network cannot express a written constraint directly and has to discover it through repeated parameter trials, so it needed extra evaluations and outer-loop architecture changes to catch up with and slightly pass P2P. Their division of labour: P2P is efficient when the key constraints are easy to state in words; P2O is more systematic when many continuous parameters need tuning.^[S1]

The best P2P policy charges hard while charge and voltage are low, lowers the current as the cell nears the high-charge, high-voltage region, and suppresses it as terminal voltage approaches 4.20 V. The authors attribute at least part of the health gain to moving charging effort away from the degradation-prone high-voltage region, while noting that the simulations do not separate the contribution of each ageing mechanism.^[S1]

In case two the authors computed, with PyBaMM, the reference current that holds heat at exactly 0.4 W (not used during optimization), and the optimized current closely matched it; P2P optimized effectively within about 150 trials, fastest in the first 50, landing between SAASBO and random search. They note that such a check against a known answer is impossible in real experiments or more complex simulations.^[S1]

The authors sketch two routes to real cells: for P2O, pick promising network architectures in simulation and run only the inner tuning loop experimentally, perhaps with multi-fidelity Bayesian optimization combining simulated and measured data; for P2P, use simulation to update the policy pool so real experiments start from a simulation-informed point. They add that the size of the gain may change under measurement noise and model mismatch, and that deployment on real cells needs model calibration, safety screening and cell-level validation.^[S1]

They also draw a boundary: the method complements surrogate modelling and differentiable optimal control, which may be better choices when a differentiable simulator, a calibrated reward model or many safe interactions are available. The paper does not compare against reinforcement learning or model-predictive control.^[S1]

What the baseline is, and what happened when it was made richer.

The baseline follows Attia and colleagues' multi-step constant-current method: three constant-current segments switching at 20%, 40% and 60% charge, also reaching 90% in 30 minutes. It has three parameters, less expressive than the language-model policies. The authors raised it to 35 segments to give it comparable degrees of freedom, but within the same budget this high-dimensional piecewise protocol was hard to optimize and its health trajectories showed pronounced sawtooth patterns. Their conclusion: too few segments restrict the solution space, too many are hard to tune without added smoothness constraints or specialised optimizers.^[S1]

How the ageing was sped up.

The cell is parameterised from the Ai2020 dataset. To see ageing within a feasible runtime, the authors scaled the SEI reaction exchange current density by five and adjusted electrode loss terms, particle cracking rates, the heat transfer coefficient and ambient temperature. To check that this keeps the ranking of protocols, they simulated three representative protocols for 1,000 cycles without acceleration, and the order held. Because the model does not include every harmful side reaction, such as electrolyte oxidation and transition-metal dissolution, they added a voltage penalty term.^[S1]

Another paper from the same group adds a caution

Ge Lei and Samuel Cooper's May 2026 preprint, Elicitation Matters, found that when a language model serves as a surrogate, the prompt's wording and query format change its predictions and confidence and so the experiments chosen. For this fast-charging paper, that means P2P's result cannot be separated from how its prompt was written; another wording may not reproduce it.^[S3]

There is a company behind this line

Samuel Cooper is Imperial's Professor of AI for Materials Design and co-founder and chief scientist of Polaron, spun out of his group in 2023; Ge Lei has worked part-time at Polaron as an LLM engineer since June 2025. The paper itself does not mention Polaron or say how the method is used there.^[S4][S5]