Skip to Content

Beyond parabolic weld bead models: AI-based 3D reconstruction of weld beads under transient conditions in wire-arc additive manufacturing

This study presents AIBead, an AI framework designed to predict and reconstruct a 3D geometry of weld beads in wire-arc additive manufacturing, especially under variable conditions where standard parabolic models are inadequate. It tackles the challenge of precise weld bead geometry prediction, offering an improvement by accurately depicting asymmetrical shapes along curved paths, thereby refining automated process planning and parameter optimization.


Type of Work: Scientific Publication

Main Author: Jan Petrik

Affiliation: ETH Zurich

Co-Authors: Benjamin Sydow, Markus Bambach

Date: 9th December 2021

Journal: Journal of Materials Processing Technology

Online: ScienceDirect

The research methodology involves the creation of a dataset using a GEFERTEC 3D metal printer and subsequent 3D scanning using an ATOS scanner to model weld beads. The AIBead model uses either Gated Recurrent Units (GRUs) or Multilayer Perceptrons (MLPs) to predict bead geometry based on G-code paths. In particular, it takes into account the effect of path curvature on bead geometry, an advance over traditional models that assume a parabolic bead shape everywhere along the path. The key is to parameterise the deposition path so that the AI algorithm has all the necessary information to make the prediction, but not an excessive amount.
The results show that AIBead performs well in predicting the asymmetric geometry of weld beads along non-straight paths, a common scenario in wire-arc additive manufacturing. Furthermore, the results highlight the important role of path curvature in geometry prediction and validate the effectiveness of the AI-based method in modelling complex geometric variations. This contribution not only advances additive manufacturing by improving the accuracy of weld bead geometry modelling, but also opens avenues for further research into adaptive process planning and control strategies.


  • Image 1
    The AIBead architecture employs MLP or GRU to model the weld bead's center line with four parameters: plane angle, Euler distance, radius, and welding speed. It predicts the 2D point cloud of the weld bead, compares it to actual data using MSE loss for optimization, and enables 3D model reconstruction from these point clouds.
  • Analysis of experimental data examines the geometric characteristics of weld bead cross-sections through centerline cuts, visualizing features such as area, toe angles, height, width, and circumference. It also graphically represents variations in these properties along the centerline, noting significant changes across the radius.
  • 3D reconstruction of a weld bead based on the deposition path. This was done using the developed machine learning architecture called AIBead.