SDG driven carbon aware machine learning model recommendation framework

Area of Science:

• Artificial Intelligence
• Machine Learning
• Environmental Sustainability

Background:

• The increasing computational demands of AI raise concerns about its carbon footprint and energy consumption.
• Sustainable AI development requires methods to balance model performance with environmental impact.

Purpose of the Study:

• To present a framework (CAOMLF) for balancing machine learning performance with environmental sustainability.
• To provide data-driven recommendations for AI models that minimize carbon emissions and energy usage.

Main Methods:

• Developed a multi-objective framework (CAOMLF) analyzing carbon emissions and energy consumption.
• Evaluated basic machine learning tasks (classification, regression, clustering) across diverse data types (text, image, tabular).
• Systematically assessed five distinct model architectures for each task, capturing accuracy and carbon emission data.

Main Results:

• The framework generates a trade-off score to balance predictive performance and carbon footprint.
• Recommendations are provided for optimal hyperparameters and model selection based on the trade-off score.
• The 'train once' approach in the recommendation phase suggests efficient models for performance and environmental impact.

Conclusions:

• CAOMLF offers a systematic approach to sustainable AI development.
• The framework aids in selecting machine learning models that are both high-performing and environmentally conscious.
• This research contributes to reducing the environmental impact of AI through informed model selection.