PCF-LLM: scaling LLMs for multimodal understanding of structured scientific data in photonic crystal fiber sensors

• We define the task of unified PCF understanding, integrating four key sub-tasks (property prediction, inverse design, structural description, property interpretation) into a single multimodal learning framework. • We propose PCF-LLM , the first architecture-agnostic LLM framework that jointly models PCF geometry, optical properties, and textual descriptions via Cross-Modality Alignment, enabling multi-task reasoning wihout retraining for each objective. • We curate PCF-MM-170K , the first large-scale multimodal PCF dataset, comprising 170K structured samples across four representative architectures, paired with fine-grained simulation and text annotations. The dataset will be publicly released to support future research. • We benchmark PCF-LLM on diverse LLM backbones (API-based and open-sourced) and tasks, showing consistent accuracy gains, strong physical consistency, and generalization to out-of-distribution structures. Photonic crystal fibers (PCFs) exhibit complex and highly tunable structure–property relationships, making them promising for diverse photonic applications but challenging for accurate modeling and inverse design. Traditional numerical solvers offer high fidelity but are computationally expensive, while existing learning-based approaches are typically limited to narrow, single-task objectives and generalize poorly to unseen structures. We define PCF understanding as the capability to jointly reason over numerical PCF geometry–property mappings and textual descriptions, enabling four core tasks: optical property prediction, inverse design suggestion, structural description generation, and property interpretation. To address these, we propose PCF-LLM, a scalable multimodal framework that adapts pretrained large language models (LLMs) for unified PCF understanding. PCF-LLM incorporates a cross-modality alignment mechanism to fuse structured PCF geometry and optical properties with language prompts, and employs parameter-efficient fine-tuning via Low-Rank Adaptation. To enable such modeling, we curate PCF-MM-170K, the first large-scale multimodal PCF dataset comprising 170,000 samples across four representative structures, each annotated with high-fidelity optical simulations and fine-grained textual descriptions. Extensive experiments across multiple LLMs demonstrate that PCF-LLM achieves high accuracy, strong physical consistency, and robust cross-task generalization, advancing the use of LLMs for scientific discovery in photonics.