CAS-NAS: A carbon-aware neural architecture search framework for sustainable AI development

The exponential growth of artificial intelligence has precipitated significant environmental concerns, with large-scale models consuming massive computational resources and generating substantial carbon emissions. This research addresses this critical challenge through a novel carbon-aware neural architecture search (CAS-NAS) framework that optimizes deep learning models across three objectives: predictive accuracy, energy efficiency, and carbon footprint. Our NSGA-II-based optimization demonstrates 30%–42% reductions in energy consumption (Joules/inference) and 28%–38% lower carbon emissions (gCO 2 e) compared to ResNet-50 and EfficientNet baselines, while maintaining competitive accuracy (less than 2% drop). The methodology integrates real-time carbon intensity simulation with hardware-aware energy estimation and multi-objective evolutionary algorithms to discover Pareto-optimal architectures. Our comprehensive evaluation system uses proxy datasets (CIFAR-10/100) and hardware profiles (Raspberry Pi 4, NVIDIA Jetson) to simulate real-world conditions without physical data collection. The framework incorporates sparsity-driven quantization, dynamic precision scaling, and carbon-aware scheduling to adapt models to variable grid conditions. Experimental results reveal critical trade-offs: a 1% accuracy reduction yields 45%–60% energy savings, while temporal carbon shifting decreases emissions by 40% during renewable energy availability. Our work establishes the first standardized evaluation metrics for Green Artificial Intelligence (Green AI): Carbon Efficiency Score (accuracy per gCO 2 e) and Energy-Accuracy Ratio enabling direct comparison of sustainable algorithms. This research provides both a methodological blueprint for environmentally conscious AI development and empirical evidence that substantial emission reductions are achievable through algorithmic optimization, supporting global climate initiatives like the EU Green Deal. The implementation is publicly available to accelerate adoption in edge computing and cloud infrastructure.