TCG-LLM: Physics-aware Fine-tuning with Quality-based Reward Shaping for Tropical Cyclogenesis Detection and Localization

Abstract

Tropical cyclones (TCs) are destructive weather systems, and accurate detection and localization of tropical cyclogenesis (TCG) are critical for mitigating their impacts. Existing deep learning methods mainly focus on a single TC or a limited region, and thus cannot detect and localize all concurrent TCG events over an entire ocean basin. Vision–Language Models (VLMs) can utilize multimodal inputs with excellent reasoning beyond simple perception ability for multi-factor causal reasoning problems, but they may fail to explicitly capture deep meteorological semantics without physics-aware visual encoders. To address these limitations, we propose the Tropical CycloGenesis Large Language Model (TCG-LLM), a multimodal large language model for basin-scale TCG detection and localization. We design a physics-aware fine-tuning strategy, using encoders based on convolutional neural networks to extract TC-specific physically meaningful features, which are injected in VLMs to guide the model to incorporate meteorological knowledge in fine-tuning. To improve generalization beyond training patterns, we further incorporate Group Relative Policy Optimization (GRPO) with quality-based reward shaping, using an online-learned quality function to provide dense intermediate learning signals that accelerate convergence and improve accuracy. We also apply fine-grained penalty on false negative cases to reduce missed TCG detections. We construct the Tropical Cyclogenesis Detection and Location Dataset (TCDLD) for evaluation. Experiments show that TCG-LLM reduces TC detection mean absolute error and localization mean distance error by more than 30% and 15%, respectively, compared with state-of-the-art baselines.