The rapidly growing scale and environmental cost of artificial intelligence have become criticalconcerns. Training large language models generates as much CO₂ as the lifetime of fivevehicles. However, such powerful models are not needed for all tasks. This study presents anovel framework for carbon-efficient model selection based on dataset complexity. Sevendifferent architectures—three deep learning-based models (MobileNetV3, ResNet18,EfficientNetB0), three machine learning models (Logistic Regression, Random Forest,XGBoost), and one transformer-based model (ViT-Tiny)—are comprehensively evaluated onnine image classification datasets. The dataset complexities are measured by developing ascoring system that combines metrics such as inter-class similarity, texture complexity, edgedensity, intra-class variance, and spatial frequency. This score has been shown to exhibit astrong correlation (r = 0.89) with the optimal model selection. Experimental results show thatfor simple tasks, machine learning models emit 300 times less carbon than other models, withonly a 5.4% loss in accuracy. A gradient boosting regression-based recommendation systemachieved R²=1.0 in model-task matches. The findings show that for tasks with a complexitythreshold of τ0.7, results demonstrate the need fordeep learning and transformer-based models. The results demonstrate that intelligent modelselection can reduce the carbon footprint of AI by up to 99%.
Tunahan TİMUÇİN (Thu,) studied this question.