Large language models as designers for autonomous research of anisotropic polymer thermocells

Abstract

Flexible thermoelectric materials face a persistent trade-off between output power and mechanical robustness. Here, we report OmniMat, a large language model-driven framework that designs experimental protocols in a human-in-the-loop workflow without relying on structured task-specific datasets. OmniMat integrates cross-disciplinary inspiration retrieval, mechanism-level screening, and expert-guided refinement to propose constant-velocity directional freezing in ethanol followed by salting-out reinforcement. This strategy creates continuous anisotropic ion channels within a strengthened poly(vinyl alcohol) (PVA) network. The representative 10 wt % medium-molecular-weight PVA (m-PVA) thermocell used for device validation delivers a normalized power density of 0.68 mW m − 2 K − 2 and toughness exceeding 21 MJ m − 3 . A flexible array maintains electrical output during stretching, bending, and partial cutting. These results establish a traceable route for language-model-assisted materials discovery in data-scarce research.