Hybrid NARX-LLM Model Improves Greenland Iceberg Discharge Prediction

This research introduces a Hybrid NARX-LLM framework designed to improve predictions of Greenland iceberg discharge, a complex phenomenon with limited observability. The framework integrates a traditional nonlinear autoregressive model with exogenous inputs (NARX) to capture intrinsic temporal dependencies, with a large language model (LLM) responsible for residual correction. A key innovation is the Physics-Informed Prompt (PIP) method, which translates unstructured physical knowledge into structured prompts. These prompts enable the LLM to perform zero-shot in-context reasoning, allowing it to encode glacier dynamics and environmental drivers, perceive trend patterns, and correct systematic prediction errors from the NARX component. The primary objective is to demonstrate the corrective potential of this hybrid approach, particularly in addressing extreme events and nonstationary trends that challenge traditional models. By fusing structured time-series modeling with knowledge-driven AI, the framework offers a scalable and interpretable pathway for climate forecasting, bridging data-limited scenarios with physics-informed LLM reasoning.