This is the repository for our paper "Physics-embedded graph network for accelerating phase-field simulation of microstructure evolution in additive manufacturing" on npj Computational Materials - Nature. We implemented both classic direct numerical simulation (DNS) based on the finite difference method and a new physics-embedded graph network (PEGN) approach. PEGN is a computationally light alternative for DNS. The code runs on both CPU and GPU. The DNS approach generally follows Yang et al. This code is developed under AMPL at Northwestern University.
We use JAX for implementation of the computationally intensive part. The graph construction is based on Jraph. The polycrystal structure is generated with Neper.
The typical workflow contains two major steps:
- Generate a polycrystal structure and mesh it (with Neper)
- Perform phase-field simulation.
The file src/example.py is an instructive example. To run this file, under root directory and run
python -m src.example
Please see the comments in src/example.py for further details. We also have an instruction file for FAQs.
Single-layer single-track powder bed fusion process
The left column shows results of DNS, while the right column shows results of PEGN. The first row shows temperature field, the second row is melt pool, and the the third row is grain evolution.
Multi-layer multi-track powder bed fusion process
Below is the result of using PEGN for simulating a 20-layer process.
Directional solidification
Competitive grain growth is observed if grain anisotropy is considered. Left is when grain anisotropy is NOT included, and right is when it is included. Competitiveness: red > blue > green for anisotropic case. The result is based on DNS.
