LLM-interpreted expert knowledge speeds multi-fidelity Bayesian optimization in battery design test
A new arXiv paper describes a way to feed qualitative expert knowledge about missing physics into multi-fidelity Bayesian optimization, reporting faster design search on a lithium-ion battery problem.
In plain words
Engineers often pair cheap approximate simulations with a few expensive, accurate ones to find good designs. The paper's method lets a large language model translate an expert's plain description of what the cheap simulation misses into a correction that guides which designs to test next. In a battery design test, this cut the reported search cost measure by 61% versus the conventional approach.
What was announced
Multi-fidelity Bayesian optimization (MFBO) lowers engineering design cost by combining inexpensive low-fidelity models with a small number of accurate high-fidelity evaluations, choosing designs to evaluate sequentially. The authors note a weakness: when high-fidelity data are scarce, the gap between the two model types is hard to learn from numbers alone, particularly when physics omitted from the cheap model varies with the design. Experts, however, can often state which physics is missing and when its effect grows, and existing multi-fidelity models cannot use such qualitative input directly.
The proposed LLM-Assisted MFBO adds a knowledge basis to the standard model. From an expert's description of the omitted physics, a large language model judges which of two designs needs the larger signed residual correction, and a pairwise Gaussian process turns those judgments into the basis. High-fidelity data calibrate the contribution and sign of this knowledge, so it is not trusted blindly. On a lithium-ion battery design problem, the method reportedly reduced the regret area under the curve of conventional MFBO by 61%. A controlled study that varied knowledge quality and LLM choice independently found knowledge quality, not the model choice, determined whether the search accelerated; a knowledge-level ablation quantified each statement's contribution.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, experimental setup and results have not been reviewed.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.