Friday 07 March 2025
The quest for precision in manufacturing has led researchers to develop innovative solutions that merge artificial intelligence, machine learning, and real-time control. A recent study published in a leading scientific journal takes this pursuit one step further by introducing a novel framework that leverages time-series deep neural networks to predict future states of a manufacturing process.
Directed Energy Deposition (DED), a type of additive manufacturing, is notoriously difficult to control due to its nonlinear behavior and complex interactions between the laser beam, powder bed, and molten pool. To mitigate this challenge, researchers have developed model predictive control (MPC) strategies that rely on accurate predictions of future states. However, traditional MPC approaches often struggle to capture the intricate dynamics of DED processes, leading to suboptimal performance.
Enter the Time-Series Dense Encoder (TiDE), a deep neural network designed specifically for predicting future states in DED processes. By leveraging past and future covariates, TiDE can accurately forecast melt pool temperature and depth, enabling real-time optimization of the manufacturing process. In other words, TiDE serves as a digital twin that simulates the behavior of the physical system, allowing researchers to test different scenarios and adjust parameters before actual production.
The proposed framework combines TiDE with MPC to create a robust control strategy. By feeding the predicted states into an MPC algorithm, the system can optimize the laser power, scanning rate, and other process variables in real-time. This approach not only improves the quality of the final product but also reduces porosity defects and ensures consistent melt pool depth.
The authors validated their framework using a single-track multi-layer square geometry, demonstrating its effectiveness in achieving precise temperature tracking while satisfying melt pool depth constraints within a targeted range (10%-30%). In comparison to traditional Proportional-Integral-Derivative (PID) controllers, the proposed MPC approach yielded smoother and less fluctuating laser power profiles with competitive or superior melt pool temperature control performance.
The implications of this research are far-reaching. By integrating machine learning and model predictive control, manufacturers can achieve higher precision, increased efficiency, and reduced waste in various industries, from aerospace to healthcare. As the demand for customized products grows, the ability to optimize complex manufacturing processes will become increasingly crucial.
Moreover, this study showcases the potential of artificial intelligence in transforming traditional manufacturing techniques. By embracing AI-powered digital twins and real-time optimization, manufacturers can adapt to changing market demands, improve product quality, and reduce production costs.
Cite this article: “Predictive Control in Additive Manufacturing: A Novel Framework Combining Time-Series Deep Neural Networks and Model Predictive Control”, The Science Archive, 2025.
Manufacturing, Artificial Intelligence, Machine Learning, Predictive Control, Additive Manufacturing, Directed Energy Deposition, Time-Series Deep Neural Networks, Model Predictive Control, Digital Twin, Real-Time Optimization
