A Hybrid Machine Learning and Swarm Optimization Framework for Real-Time Energy Prediction and Optimization in Additive Manufacturing

Additive Manufacturing (AM) enables unprecedented design flexibility and decentralized production; however, its high energy consumption remains a critical sustainability challenge. While recent digital twin and machine learning (ML)-based studies have focused primarily on quality monitoring, defect detection, or thermal modeling, limited research has developed integrated predictive–optimization pipelines specifically targeting real-time energy management across multiple AM platforms. This study proposes a hybrid machine learning framework that combines predictive modeling, clustering-based data compression, and swarm intelligence optimization to estimate and minimize energy consumption using real-time process parameters. Experimental data were collected from Selective Laser Sintering (SLS) and Fused Deposition Modeling (FDM) systems, incorporating layer-wise geometric descriptors, process variables (laser power, scan speed, build temperature, extrusion rate), and environmental conditions. Multiple supervised learning models—including Support Vector Regression (SVR), XGBoost, Deep Neural Networks (DNN), and a DBSCAN–XGBoost hybrid—were trained and evaluated under a controlled benchmarking protocol. The DNN with feature selection achieved the best predictive performance (RMSE = 0.72 kWh, R² = 0.96), outperforming conventional regression and ensemble models across identical datasets. Statistical comparisons confirmed a consistent improvement in prediction accuracy relative to baseline linear and non-deep learning models. The predicted energy profiles were integrated into a Particle Swarm Optimization (PSO) module to identify energy-efficient process parameter configurations under manufacturability constraints. The optimization stage achieved up to an 18% reduction in energy consumption for both SLS and FDM builds while maintaining tensile strength and dimensional accuracy within acceptable tolerances. By explicitly integrating clustering-based compression, deep learning prediction, and swarm-based optimization into a unified pipeline, this framework extends beyond existing standalone digital twin or energy estimation approaches. The results demonstrate the potential of hybrid ML–optimization architectures for scalable, real-time, energy-aware additive manufacturing and sustainable industrial deployment.