Napolitano, Enea Vincenzo (2026) Measuring and Mitigating the Environmental Impact of Artificial Intelligence. [Tesi di dottorato]

Documento PDF Napolitano_EneaVincenzo_38.pdf Download (4MB)

Collection description

Artificial Intelligence (AI) has become a transformative driver of technological and social progress, yet its rapid expansion has also revealed significant environmental challenges. The increasing complexity of AI models, coupled with massive computational requirements, has led to substantial energy consumption and greenhouse gas emissions, raising concerns about the sustainability of AI research and applications. This thesis addresses this emerging issue by proposing analytical, hybrid, and empirical frameworks aimed at measuring and mitigating the environmental impact of AI systems. The first contribution introduces an analytical metric, the AI Emission Equation, which quantifies the environmental footprint of AI tasks by integrating key variables such as computational power, energy source, training duration, and optimization strategies. This model provides a unified and interpretable approach to assess sustainability in AI experiments, validated through a real-world case study conducted at the EHBDA Research Lab. Building upon this foundation, the second contribution presents a hybrid estimation methodology that adaptively combines multiple carbon-tracking tools, such as CodeCarbon, CarbonTracker, and Green Algorithms, using a weighted aggregation system. This hybrid strategy improves the accuracy and adaptability of the estimation in different computational environments and model architectures. The third contribution offers a large-scale empirical investigation of sustainability awareness within the machine learning research community. Through a systematic analysis of peer-reviewed publications, the study reveals critical gaps in environmental reporting, optimization practices, and hardware transparency, highlighting the need for standardized sustainability protocols in AI research. The results of this work contribute both theoretical and practical insights for integrating ecological responsibility into AI development. By bridging analytical modeling, hybrid estimation, and empirical analysis, the thesis advances a framework for sustainable AI, one that balances innovation with accountability. Ultimately, this research promotes a paradigm in which environmental sustainability becomes an intrinsic measure of scientific excellence and a guiding principle for future AI innovation.

Downloads

Actions (login required)

View Item