Speaker
Description
Metal Binder Jetting (BJT-MSt/M) offers high productivity, but sintering distortion limits dimensional accuracy and first time right production. High fidelity finite element (FEM) sintering simulations can predict distortion, yet their runtime restricts design space exploration and rapid design iterations. We present a graph neural network (GNN) surrogate that predicts sintering induced deformation directly from geometry, enabling fast solver-agnostic inference.
Parts are represented as attributed graphs built from mesh connectivity and local geometric descriptors. A message passing network maps the as printed state at step k to the next state k+1 and outputs node-wise displacement vectors. The model is trained on simulated BJT-MSt/M sintering trajectories and evaluated on unseen geometries. Results show close agreement between predicted and ground truth mean nodal displacement, including transient peaks, with small absolute errors in the mean displacement signal. Inference is orders of magnitude faster than FEM, enabling near real time distortion estimation and supporting iterative compensation loops.
This framework provides a scalable path toward geometry aware, data driven sintering distortion prediction and motivates future work.
| Professional Status of the Speaker | Doctoral or Master Student |
|---|---|
| Invitation letter for visa | No |
| Interest in submitting a paper in a special issue of | Advanced Engineering Materials (Wiley) |
