AI-driven process parameter reconfiguration in DED-arc
Loading...
Date
2026-07-08
Authors
Γεωργούσης, Ευάγγελος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Modern manufacturing is undergoing a drastic transformation driven by Artificial Intelligence (AI) and advanced sensor systems, enabling smarter, autonomous and flexible production processes across industries. Simultaneously, Additive Manufacturing (AM) is rapidly expanding as a viable alternative to conventional subtractive methods, due to its advantages in design freedom and reduced lead times and cost. DED-Arc stands out for the fabrication of metallic components due to its high deposition rates, low material cost and large-scale build capabilities. However, the process’ high complexity and sensitivity to process parameter selection makes consistent control of deposited bead geometry a persistent challenge limiting industrial adoption, pushing AI and Machine Learning (ML) based approaches to the forefront of recent literature. This study presents an end-to-end framework for bead geometry prediction and process parameter optimization in DED-Arc. Experiments were conducted using Lincoln Electric’s PGMAW process following an extended L-16 Taguchi Design of Experiments across four process parameters: Wire Feed Speed (WFS), Travel Speed (TS), Voltage (V) and Pulse Frequency (P), with bead width and height being selected as the primary response variables. Analysis of Variance (ANOVA) was applied on the experiments to determine the effects of process parameters on the response, and a GPR model with optimized hyperparameters was trained and validated for bead geometry prediction, demonstrating a strong predictive performance for bead width (R2=0.88) and bead height (R2=0.69). Furthermore, the models were integrated into a multi-objective optimization framework utilizing a Genetic Algorithm (GA), NGSA-II, generating a Pareto front of optimal process parameter combinations for target bead geometries. To further address data scarcity, a second experimental dataset with a different wire diameter was used to benchmark transfer learning algorithms, TrAdaBoost and Two-Stage TrAdaBoost, against source-only and pooled-data baselines. Both algorithms demonstrated reliable predictive performance with R²>0.9 using as few as 6 experimental points for bead width and R2>0.83 using 10 experimental points for bead height, confirming practical viability for rapid reconfiguration with minimal experimentation.
Description
Keywords
Directed energy deposition (DED), Machine learning, Transfer learning, Process parameter optimization, Rapid reconfiguration