NanoConfinementChat: A Domain-Specific Large Language Model for Mechanistic Interpretation of Nanoconfined Water and Ion Transport in Membrane Nanochannels

Nanoconfined water and ion transport determines permeability and selectivity in membrane nanochannels, yet the relevant mechanisms are dispersed across material systems, molecular simulations, and theoretical studies, hindering systematic mechanism-oriented knowledge synthesis. Here, we develop NanoConfinementChat, a domain-specific large language model (LLM) for analyzing transport mechanisms in ion-selective membrane nanochannels. A curated corpus of 518 papers was used to construct 8500 mechanism-oriented instruction–answer pairs for supervised fine-tuning (SFT) of Llama-3.1-8B-Instruct using low-rank adaptation (LoRA). A retrieval-augmented generation (RAG) module was further integrated for literature grounding and source tracing, and model performance was assessed using a 90-question gold-point benchmark. Without RAG, NanoConfinementChat achieved an average benchmark score of 2.144 and a major scientific error rate of 2.22%, compared with 11.11–13.33% for the two baseline models. Case studies showed improved mechanistic interpretation of ordered water transport, dehydration-driven ion exclusion, graphene oxide interlayer-spacing effects, electrostatic exclusion, and evidence requirements for interfacial slip. Domain fine-tuning provided the primary performance gain, while RAG improved source traceability. The framework offers a lightweight approach for evidence-grounded interpretation of structure–mechanism–transport relationships in membrane nanochannels.