Green-NAS: A Global-Scale Multi-Objective Neural Architecture Search for Robust and Efficient Edge-Native Weather Forecasting

TL;DR

Green-NAS presents a multi-objective neural architecture search framework tailored for edge-native weather forecasting, prioritizing accuracy and efficiency to reduce energy use and carbon footprint. By employing NSGA-II across a diverse temporal search space, the approach discovers a Pareto front of architectures, including a high-accuracy GRU-based model (Green-NAS-A) and ultra-compact CNNs (Green-NAS-C) suitable for IoT devices. The study demonstrates strong parameter efficiency (up to 35,500x fewer parameters than GraphCast) with competitive RMSE, robust uncertainty quantification via conformal prediction, and sub-millisecond edge inference. Transfer learning further boosts performance by about 5.2% under full data and remains advantageous with limited data, enabling rapid, scalable deployment in data-sparse regions like the Global South. Collectively, Green-NAS showcases a path toward equitable, environmentally sustainable climate AI for global-scale, edge-enabled forecasting.

Abstract

We introduce Green-NAS, a multi-objective NAS (neural architecture search) framework designed for low-resource environments using weather forecasting as a case study. By adhering to 'Green AI' principles, the framework explicitly minimizes computational energy costs and carbon footprints, prioritizing sustainable deployment over raw computational scale. The Green-NAS architecture search method is optimized for both model accuracy and efficiency to find lightweight models with high accuracy and very few model parameters; this is accomplished through an optimization process that simultaneously optimizes multiple objectives. Our best-performing model, Green-NAS-A, achieved an RMSE of 0.0988 (i.e., within 1.4% of our manually tuned baseline) using only 153k model parameters, which is 239 times fewer than other globally applied weather forecasting models, such as GraphCast. In addition, we also describe how the use of transfer learning will improve the weather forecasting accuracy by approximately 5.2%, in comparison to a naive approach of training a new model for each city, when there is limited historical weather data available for that city.

Figures (3)

• Figure 1: Pareto Front of Discovered Architectures. The figure illustrates the tradeoff between forecasting performance (RMSE on y-axis) and model complexity (number of parameters on x-axis in a log scale), for all 20 models that were discovered by NAS. Green-NAS-A/B/C (stars colored green) are superior to those found with random search and superior to several hand-tuned baselines, each corresponding to an efficient accuracy point.
• Figure 2: Transfer Learning Efficiency. This figure illustrates how much more effective transfer learning is than training from scratch regardless of the amount of data used (1%, 10%, 50%, or 100%). Transfer learning (green) always performs better than training from scratch (red). The standard deviation (N = 10 trials) is represented by the error bars. Even with 100% of the data, there is a statistically significant increase in accuracy of +5.2% ($p < 10^{-12}$) for pre-training, which shows that pre-training successfully transfers global weather knowledge into the target cities.
• Figure 3: Feature Importance (Permutation). Relative importance of input features for Green-NAS-A predictions. Temperature and Pressure dominate, but the model utilizes all available variables.