Green metrics for AI: A hybrid strategy for environmental impact assessment

Accurately assessing the environmental impact of Artificial Intelligence (AI) systems is essential to promoting sustainable development within the field. However, existing carbon emission tracking tools often rely on divergent assumptions, limited hardware compatibility, and inconsistent measurement granularity, leading to fragmented and often incomparable results. In this paper, we propose a hybrid methodology that integrates a weighted aggregation of ten widely-used emission estimation tools with a dynamic model-to-tool matching procedure. The weighted strategy assigns normalized importance scores to each tool based on six criteria. In parallel, the dynamic matching component identifies the most context-appropriate estimation tool for a given AI model. We evaluate our methodology on multiple architectures, including MobileNet, DenseNet, and GPT-3, deployed in different computing environments. Results demonstrate that our combined approach enhances both estimation accuracy and interpretability, outperforming single-tool assessments and offering a scalable solution for transparent carbon accounting in AI research.