Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Virtual Foundry Graphnet for Metal Sintering Deformation Prediction

Rachel, Chen, Juheon Lee, Chuang Gan, Zijiang Yang, Mohammad Amin Nabian, Jun Zeng

TL;DR

The paper introduces Virtual Foundry Graphnet, a graph neural network–based simulator to accelerate metal sintering deformation predictions for HP's Digital Twin. By voxelizing parts into graphs and applying an encoder–processor–decoder with multi-step, physics-informed loss, the model predicts deformation states and rolls out future states with near real-time performance. Empirical results on multiple geometries show millimeter- to sub-millimeter-scale accuracy and runtimes around one minute for full simulations, representing a substantial speed-up over traditional FE approaches. The work is open-sourced on the NVIDIA Modulus platform to foster community-driven development and scaling to diverse geometries and processing configurations.

Abstract

Metal Sintering is a necessary step for Metal Injection Molded parts and binder jet such as HP's metal 3D printer. The metal sintering process introduces large deformation varying from 25 to 50% depending on the green part porosity. In this paper, we use a graph-based deep learning approach to predict the part deformation, which can speed up the deformation simulation substantially at the voxel level. Running a well-trained Metal Sintering inferencing engine only takes a range of seconds to obtain the final sintering deformation value. The tested accuracy on example complex geometry achieves 0.7um mean deviation for a 63mm testing part.

Virtual Foundry Graphnet for Metal Sintering Deformation Prediction

TL;DR

The paper introduces Virtual Foundry Graphnet, a graph neural network–based simulator to accelerate metal sintering deformation predictions for HP's Digital Twin. By voxelizing parts into graphs and applying an encoder–processor–decoder with multi-step, physics-informed loss, the model predicts deformation states and rolls out future states with near real-time performance. Empirical results on multiple geometries show millimeter- to sub-millimeter-scale accuracy and runtimes around one minute for full simulations, representing a substantial speed-up over traditional FE approaches. The work is open-sourced on the NVIDIA Modulus platform to foster community-driven development and scaling to diverse geometries and processing configurations.

Abstract

Metal Sintering is a necessary step for Metal Injection Molded parts and binder jet such as HP's metal 3D printer. The metal sintering process introduces large deformation varying from 25 to 50% depending on the green part porosity. In this paper, we use a graph-based deep learning approach to predict the part deformation, which can speed up the deformation simulation substantially at the voxel level. Running a well-trained Metal Sintering inferencing engine only takes a range of seconds to obtain the final sintering deformation value. The tested accuracy on example complex geometry achieves 0.7um mean deviation for a 63mm testing part.
Paper Structure (15 sections, 2 equations, 10 figures, 2 tables)

This paper contains 15 sections, 2 equations, 10 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Graph neural networks
  4. Digital Twin
  5. Sintering physics
  6. Physics informed network
  7. The proposed method
  8. Data preparation
  9. Model architecture and training
  10. Objective functions
  11. Inferencing
  12. Experimental results
  13. Model optimization
  14. Test geometry accuracy
  15. CONCLUSION AND FUTURE WORK

Figures (10)

  • Figure 1: Metal Jet print stages to obtain the final part.
  • Figure 2: The Stanford dragon test model shows the isotropic part shrinkage, including gravitational sag, slump, and bending.
  • Figure 3: Data process pipeline to obtain the sintering-time graph data for training and inferencing. The sample dragon build is voxelated, the color gradient shows each voxel’s deformation scale; we preprocess the voxel-based simulation output (left) to form “nodes” of the graph (right), representing the concept of “metal particles”.
  • Figure 4: Data process pipeline to obtain the sintering-time graph data for training and inferencing. The preprocessing steps are to convert the voxel-based simulation data into a graph-based data structure for each sintering step
  • Figure 5: Simulation prediction on the processed graph data with graph neural network architecture
  • ...and 5 more figures