Effective job reassignments in large scale collaborative additive manufacturing networks

Abstract

The growth of large-scale collaborative additive manufacturing (AM) networks necessitates scalable, efficient, and privacy-preserving solutions for decentralized production planning. This study investigates the integration of machine learning (ML) into combinatorial reverse auction frameworks to support cost-efficient job reassignments across distributed AM systems. We benchmark several supervised ML models trained on optimal solutions to a single-machine AM scheduling problem and identify robust, regularised linear regression models as the best-performing predictors. Our best model achieves a mean absolute percentage error of approximately 3%, allowing for rapid and reliable cost predictions. Our experiments further demonstrate that linear regression models can outperform more complex alternatives such as neural networks and decision tree ensembles in both accuracy and robustness. The ML-enhanced framework significantly reduces computational overhead and limits the exposure of sensitive production data, outperforming traditional approaches like mixed-integer linear programming and Adaptive Large Neighborhood Search. When integrated into a decentralised auction mechanism, the model enables efficient task reallocation and system-wide cost reductions. While occasional violations of individual rationality due to cost underestimation present a drawback compared to benchmark methods, we argue that long-term efficiency gains may offset these effects in repeated interactions. Overall, this work highlights the potential of lightweight ML models to enable scalable, adaptive, and privacy-aware coordination in decentralised AM networks.