Carbon-Aware Training Schedules for Machine Learning Models: An Energy-Efficient Green AI Approach

The proliferation of machine learning (ML) applications has drastically raised the energy demand of computations, which inevitably contributes to CO₂ emissions, posing serious threats to the sustainability of AI models. Although existing studies have made tremendous efforts to improve the efficiency of models with respect to accuracy, little work has been dedicated to exploring the sustainable impact of model training. To address the limitations in existing models, this paper proposes the usage of carbon-aware training schedules as an efficient Green AI technique that helps to significantly decrease the energy demand as well as CO2 emissions associated with machine learning models without impairing their prediction performance. To evaluate our proposed technique, we experiment with various ML models with different training scenarios, including different values of the batch size, number of epochs, strategies, and early stopping. The energy requirements of various trained models are estimated with reference to their training time and power models, and CO2 emissions are established with reference to the known emission factors. Experimental studies on familiar benchmark data teaching schedules demonstrate that, when equipped with carbon-aware training schedules, significant savings in energy consumption and CO2 emissions can be made without adversely affecting the predictive accuracy. The paper will be useful to the existing body of research regarding Green AI due to its simplicity as a model-agnostic approach to incorporating carbon efficiency into machine learning training procedures.