SUSTAINABLE AI: REDUCING ENERGY CONSUMPTION IN MACHINE LEARNING MODEL

Abstract

Abstract: The rapid growth of large machine learning (ML) models has led to significant increases in energy consumption and carbon emissions, which has raise significant environmental concerns. This work investigates a novel technique designed to increase energy efficiency across different stages of machine learning development. It examines literature from 2018 to 2025 and investigates energy-efficient techniques. The study spans across 6 model-compression techniques namely as pruning, quantization, knowledge distillation, hardware-aware neural architecture search, efficient optimization methods, and scheduling. Renewable energy sources are also included in it. We found that aggressive quantization can reduce power use by as much as 75-85% with less than 1% accuracy loss, depending on the precision level used. Power can be conserved through structured pruning without affecting performance, on the other hand. Further to validate our study we used case study of big language models, which showed how large the problem really is and what can be done to solve it. And, based on that three key research gaps were identified: the lack of uniform energy standards, the limited implementation of efficiency improvements across different fields, and the inadequate integration of renewable energy planning into machine learning infrastructure. Based on these results, we proposed a research agenda to develop AI systems that are both environment friendly and computationally efficient. Keywords: Sustainable AI, Green AI, Energy-Efficient Machine Learning, Model Compression, Neural Architecture Search, Carbon Emissions, Quantization, Knowledge Distillation