AEON: An Energy-Aware Optimization Engine for Python Machine Learning Pipelines

Abstract

The rapid integration of machine learning (ML) into edge devices, IoT ecosystems, and cloud computing environments has significantly increased the need for energy-efficient approaches to model training and inference. Although existing tools primarily concentrate on either measuring energy consumption or optimizing deep learning workloads, a comprehensive solution capable of optimizing the energy efficiency of arbitrary Python-based machine learning pipelines remains absent. Aeon is introduced as a fully automated, Abstract Syntax Tree (AST) - level energy optimization engine that accepts Python ML code, profiles energy consumption per pipeline stage, applies a portfolio of eight orthogonal optimization strategies, and selects the Pareto-optimal configuration across energy, accuracy, and speed. The Energy–Accuracy Optimization Index (EAOI) is defined as a unified metric representing the ratio of retained accuracy to normalized energy consumption, enabling consistent cross-strategy comparison. Experimental evaluation across four representative machine learning workloads - linear regression, random forest classification, multilayer perceptron, and object detection shows energy savings ranging from 34% to 49%, while maintaining accuracy degradation below 0.5%, and achieving an average EAOI of 1.62. Aeon generates an optimized rewritten Python program and a detailed analytical report, making the framework immediately usable for practitioners deploying ML on resource-constrained platforms.