Virtual Foundary GraphNet
Metal sintering is a necessary step for Metal Injection Molded parts and binder jetting such as HP’s metal 3D printer (MetJet). The metal sintering process introduces large deformation varying from 25% to 50% depending on the green part porosity. The final part’s geometrical accuracy and consistency remain the top challenge to manufacturing yield. This is due to:
Green parts out of MetJet printer are much more porous than other technologies (e.g., MIM);
Such shrinkage is not isotropic depending on non-uniform stress built up during the sintering process, e.g., gravitational sag, gravitational slump, surface drag.
In this work, 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, for a single sintering step (equivalent to 8.3 minutes physical sintering time), and a 0.3mm mean deviation for the complete sintering cycle (~4 hrs physical sintering time).
Full paper on: Virtual Foundry Graphnet for Metal Sintering Deformation Prediction
For more sample parts simulation:
Download Modulus, make or install
Find the matching torch-scatter version with torch and cuda version enabled:
i.e. pip install torch-scatter-f
https://data.pyg.org/whl/torch-2.2.0%2Bcu121/torch_scatter-2.1.2%2Bpt22cu121-cp311-cp311-linux_x86_64.whl(replace the torch-scatter wheel with the matching cuda, torch version )torch-scatter installation guide:
https://pypi.org/project/torch-scatter/wheels source:
https://data.pyg.org/whl/
pip install tensorflow
test version: tensorflow-2.15.0.post1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
for logging:
pip install wandb
pip install mlflow
For training with mixed precision:
https://github.com/NVIDIA/apex
pyvista is required only if need to run data proprocessing with the raw simulation data files
Dev:
install pytest
pip install importlib-metadata
pip install hydra-core –upgrade
Change the params in conf/config.yaml for training:
mode: “train”
ckpt_path_vfgn={path to save model trained ckpt}, i.e. “models/test24”
data_path: {data path for the pre-processed data in tfrecord}, i.e. “./data/test_validation”
noise_std: i.e.1e-9
loss: i.e. me loss # options: [‘standard’, ‘anchor’, ‘me’, ‘weighted’, ’‘’correlation’, ‘anchor_me’]
Then run:
python train.py
Currently default params:
INPUT_SEQUENCE_LENGTH = 5
PREDICT_LENGTH = 1
NUM_PARTICLE_TYPES = 3
Provided ckpt trained at every 100 step of Physcis sintering simulation data
Change the params in conf/config.yaml for testing:
mode: “eval_rollout”
eval_split: “test”
batch_size: 1
noise_std: 0
ckpt_path_vfgn={path to model trained ckpt}, i.e. “models/ckpt/model_loss-4.17E-06_step-1113000.pt”
output_path: {path to store outputs}, i.e. “rollouts/test24”
data_path: {preprocessed test data tfrecord}, i.e. “./data/test_validation”
Then run:
python train.py
Change the params in conf/config.yaml:
rollout_path={selected_prediction_output.json}, i.e. “rollouts/rollout_test_0.json”
metadata_path={metadata path}, i.e. “./data/test_validation”
test_build_name={test file name}, i.e. “test0”
python render_rollout.py
Change the params in conf/config.yaml for inference run:
(model tested with spliting the entire sintering profile into 2 stages, can combine the entire sintering profile inferencing according to train schema)
mode: “rollout”
eval_split: “inference” # name of the tfrecord dataset
noise_std: 0
batch_size: 1
ckpt_path_vfgn={path to model trained ckpt}, i.e. “models/ckpt/models/ckpt/model_loss-4.17E-06_step-1113000.pt”
output_path: {path to store outputs}, i.e. “rollouts/test24”
data_path: {preprocessed test data tfrecord}, i.e. “./data/test_validation”
python inference.py
Test data
Same voxel resolution as train
To generate your own tfrecord from Physical simulation output:
python data_process/rawdata2tfrecord.py
Defition of step_context & methods tried:
appending only the previous step global context / ( sinter temperature)
tensor_dict['step_context'] =tensor_dict['step_context'][-predict_length - 1][tf.newaxis]
appending previous sequence of global context / (sequence of sinter temperature)
tensor_dict['step_context'] = tf.reshape(tensor_dict['step_context'][:-1], [1, -1])
appending the entire sequence of sintering profile
tensor_dict['step_context'] = tf.reshape(tensor_dict['step_context'],[1, -1])
With the model prediction accuracy and fast inference speed, this work, as a component of HP’s Digital Twin effort, Virtual Foundry Graphnet led by HP Labs, aims to apply Physics-ML to significantly accelerate the computation that predicts the metal powder material phase transition. It has achieved orders of magnitude speed-up compared to physics simulation software while preserving reasonable accuracy. Furthermore, Virtual Foundry Graphnet has demonstrated an outstanding path forward to scaling for diverse parts of arbitrary geometrical complexity and scaling for different process parameter configurations.
Learning to Simulate Complex Physics with Graph Networks
@inproceedings{sanchezgonzalez2020learning,
title={Learning to Simulate Complex Physics with Graph Networks},
author={Alvaro Sanchez-Gonzalez and
Jonathan Godwin and
Tobias Pfaff and
Rex Ying and
Jure Leskovec and
Peter W. Battaglia},
booktitle={International Conference on Machine Learning},
year={2020}
}