Synchronization-avoiding-algorithms

Code references for the numerical examples in the paper:

Data-driven synchronization-avoiding algorithms in the explicit distributed structural analysis of soft tissue, Guoxiang Grayson Tong and Daniele E. Schiavazzi, 2022, Computational Mechanics

Descriptions:

Mesh_info:

Folder containing gmsh source file and a unstructured cantilever mesh

Results:

Folder will contain computed FEA solution, modelled solution, and mesh partitioning information

Tools:

Code for building 3D elastodynamics solver using linear finite element, construction of LSTM-encoder-decoder network, etc.

Data_prepare.py

Main script to run the explicit finite element elastodynamics solver to gather training data.

Shared_extraction.py

Extract displacement degrees of freedom on all shared nodes, from pre-computed numerical solutions.

Model_training.py

Main training script for the LSTM-encoder-decoder network

Onliner_predictor.py

Main script to perform the proposed synchronization-avoiding algorithm, using the pre-trained network model.

Working flow:

A minimal working example is included in this repository.

• Execute mpirun -np 2 python3 Data_prepare.py. The processor-wise solutions will be saved to Results/Dynamics/.
• Execute mpirun -np 2 python3 Shared_extraction.py. The processor-wise shared degrees of freedom will be saved to Results/sol_on_shared/.
• Execute mpirun -np 2 python3 Model_training.py. The trained model, training/validation error, etc will be saved to Distributed_save/.
• Execute mpirun -np 2 python3 Online_predictor.py. The predicted solution will be saved to Results/Dynamics/.
• Execute python3 plotter.py inside Results/. The comparison picture can be seen as Comparison.pdf.

Note: The hyperparameter setting in Model_training.py is primarily for fast training, so not optimal.

Dependencies:

• MPI for distributed computation: mpi4py
• Meshing: gmsh
• Mesh processing meshio
• Mesh partitioning: mgmetis
• Neural network building: PyTorch
• Code reference for LSTM-encoder-decoder network: LSTM