Towards Affordable Weather Forecasting in South Africa: Parameter-Efficient Fine-Tuning of an AI Weather Foundation Model

The scarcity of high-quality weather data, computational resources, and technical expertise make AI weather prediction particularly challenging across Africa. This study investigates whether recent advances in deep learning can help overcome these barriers by enabling cost-effective, accurate regional weather forecasting. We begin with a cross-regional evaluation and observe that Aurora, a state-of-the-art deep learning model for global weather forecasting, exhibits comparatively weaker performance over South Africa than over Europe and the United States. This observation motivates our fine-tuning experiment, aimed at improving the model’s performance across South Africa. We explore parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) on the smallest variant of Aurora. The model is fine-tuned specifically for the South African region using a subset of HRES T0 data. Despite being trained on a limited dataset with modest computational resources, we end up with a fine-tuned model that consistently outperforms both the pretrained small and large Aurora models across several important meteorological variables. These results highlight the potential of fine-tuning existing models to provide competitive and localized forecasts in data-scarce, resource-constrained settings.