EC4A: Physics-Preserving Compression and Regional Downscaling for Edge AI Weather Forecasting in Africa

AI-based weather models provide increasingly skillful global forecasts but remain difficult to deploy on resource-constrained infrastructure. This limitation is particularly important for African National Meteorological and Hydrological Services, where computing resources, observational coverage, and connectivity are limited. We propose EC4A, a teacher--student framework that compresses a global AI weather model into a CPU-compatible student while preserving global accuracy and enhancing regional skill. The framework combines block-wise feature distillation, structured pruning, INT8 quantization-aware training, physics-aware output bounding, and regional downscaling using high-resolution land-surface covariates. Lightweight low-rank adapters enable regional adaptation without retraining the full backbone, while federated learning supports updates without sharing raw observations. The lightweight students will be first evaluated on their forecast accuracy, extreme-event skill, physical consistency, and regional downscaling quality, and then on the inference latency, peak memory, and energy consumption. The intended climate-impact pathway is to make locally deployable weather information more accessible for agricultural planning, disaster management, and resource optimization, to ensure a more robust climate-risk adaptation in Africa.