Green-NAS (Edge-native weather forecasting)

Problem#

Many forecasting models are designed assuming abundant compute, but practical deployments (especially in low-resource settings) must operate under constraints like limited memory, low power, and latency requirements.

Approach#

Green-NAS frames architecture search as a multi-objective optimization problem:

• maintain strong predictive performance
• aggressively control model size and efficiency targets aligned with “Green AI” principles

Key result (high-level)#

The best discovered model achieves competitive forecasting accuracy while using orders of magnitude fewer parameters than large, globally deployed systems, making deployment feasible on constrained hardware.

Links#

• Paper (arXiv): https://arxiv.org/abs/2602.00240
• PDF: https://arxiv.org/pdf/2602.00240.pdf

Reproducibility checklist (what I aim to provide)#

• dataset + preprocessing notes
• baseline configuration
• search space definition
• evaluation metrics + reporting template