![Welcome to [Skardova, Daby-Seesaram & Genet, Submitted]'s demos! - Home](../_static/logo.png){kind=logo}
3D Case#
#%% Libraries import
# import HiDeNN library
import sys
# sys.path.append("../neurom/")
from neurom.HiDeNN_PDE import MeshNN, NeuROM, MeshNN_2D, MeshNN_1D, MeshNN_3D
# Import pre-processing functions
import neurom.src.Pre_processing as pre
# Import torch librairies
import torch
import torch.nn as nn
# Import mechanical functions
from neurom.src.PDE_Library import Strain, Stress,VonMises_plain_strain
# Import Training funcitons
from neurom.src.Training import Training_2_3D_FEM
#Import post processing libraries
import neurom.Post.Plots as Pplot
import time
import os
import torch._dynamo as dynamo
mps_device = torch.device("mps")
from importlib import reload # Python 3.4+
import tomllib
import numpy as numpy
import argparse
Load configuration file for 3D case
Default_config_file = 'Configurations/config_3D_Cube_w_hole.toml'
with open(Default_config_file, mode="rb") as f:
config = tomllib.load(f)
Experiment 1#
fixed mesh
multi-level training
optimizers:
l-bfgs
Adam + l-bfgs
element_size = 1
max_refinment = 5
config["solver"]["FrozenMesh"] = True
config["interpolation"]["MaxElemSize2D"] = element_size
optimizers = ["adam","lbfgs"]
for o in optimizers:
config["training"]["optimizer"] = o
for refin in range(1,max_refinment+1):
config["training"]["multiscl_max_refinment"] = refin
if config["interpolation"]["dimension"] == 1:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["geometry"]["A"] # Section area of the 1D bar
)
elif config["interpolation"]["dimension"] in [2, 3]:
try:
Mat = pre.Material( flag_lame = False, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["material"]["nu"] # Poisson's ratio
)
except:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["lmbda"], # First Lame's coef
coef2 = config["material"]["mu"] # Second Lame's coef
)
MaxElemSize = pre.ElementSize(
dimension = config["interpolation"]["dimension"],
L = config["geometry"]["L"],
order = config["interpolation"]["order"],
np = config["interpolation"]["np"],
MaxElemSize2D = config["interpolation"]["MaxElemSize2D"]
)
Excluded = []
Mesh_object = pre.Mesh(
config["geometry"]["Name"], # Create the mesh object
MaxElemSize,
config["interpolation"]["order"],
config["interpolation"]["dimension"]
)
Mesh_object.AddBorders(config["Borders"]["Borders"])
Mesh_object.AddBCs( # Include Boundary physical domains infos (BCs+volume)
config["geometry"]["Volume_element"],
Excluded,
config["DirichletDictionryList"]
)
Mesh_object.MeshGeo() # Mesh the .geo file if .msh does not exist
Mesh_object.ReadMesh() # Parse the .msh file
Mesh_object.ExportMeshVtk()
if int(Mesh_object.dim) != int(Mesh_object.dimension):
raise ValueError("The dimension of the provided geometry does not match the job dimension")
Model_FEM = MeshNN_3D(Mesh_object, n_components = 3)
Model_FEM.Freeze_Mesh()
Model_FEM.UnFreeze_FEM()
if not config["solver"]["FrozenMesh"]:
Model_FEM.UnFreeze_Mesh()
Model_FEM.RefinementParameters( MaxGeneration = config["training"]["h_adapt_MaxGeneration"],
Jacobian_threshold = config["training"]["h_adapt_J_thrshld"])
Model_FEM.TrainingParameters( loss_decrease_c = config["training"]["loss_decrease_c"],
Max_epochs = config["training"]["n_epochs"],
learning_rate = config["training"]["learning_rate"])
Model_FEM.train()
Model_FEM = Training_2_3D_FEM(Model_FEM, config, Mat)
training_time = Model_FEM.training_recap["Duration_tot"]
torch.save(Model_FEM.state_dict(), 'Trained_models/Experiment_1/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item()))
numpy.save('Trained_models/Experiment_1/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item()) +"_time.npy", training_time)
# Pplot.ExportFinalResult_VTK(Model_FEM,Mat,config["postprocess"]["Name_export"]+"_"+str(Model_FEM.state_dict()["h_max"].item())+"_final_1")
# Pplot.ExportHistoryResult_VTK(Model_FEM,Mat,config["postprocess"]["Name_export"])
Experiment 2#
fixed mesh
single-level training
optimizers:
l-bfgs
Adam + l-bfgs
element_sizes = [1, 0.5, 0.25, 0.125, 0.0625]
config["training"]["multiscl_max_refinment"] = 1
config["solver"]["FrozenMesh"] = True
optimizers = ["lbfgs","adam"]
for o in optimizers:
config["training"]["optimizer"] = o
for e in element_sizes:
config["interpolation"]["MaxElemSize2D"] = e
if config["interpolation"]["dimension"] == 1:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["geometry"]["A"] # Section area of the 1D bar
)
elif config["interpolation"]["dimension"] in [2, 3]:
try:
Mat = pre.Material( flag_lame = False, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["material"]["nu"] # Poisson's ratio
)
except:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["lmbda"], # First Lame's coef
coef2 = config["material"]["mu"] # Second Lame's coef
)
MaxElemSize = pre.ElementSize(
dimension = config["interpolation"]["dimension"],
L = config["geometry"]["L"],
order = config["interpolation"]["order"],
np = config["interpolation"]["np"],
MaxElemSize2D = config["interpolation"]["MaxElemSize2D"]
)
Excluded = []
Mesh_object = pre.Mesh(
config["geometry"]["Name"], # Create the mesh object
MaxElemSize,
config["interpolation"]["order"],
config["interpolation"]["dimension"]
)
Mesh_object.AddBorders(config["Borders"]["Borders"])
Mesh_object.AddBCs( # Include Boundary physical domains infos (BCs+volume)
config["geometry"]["Volume_element"],
Excluded,
config["DirichletDictionryList"]
)
Mesh_object.MeshGeo() # Mesh the .geo file if .msh does not exist
Mesh_object.ReadMesh() # Parse the .msh file
Mesh_object.ExportMeshVtk()
if int(Mesh_object.dim) != int(Mesh_object.dimension):
raise ValueError("The dimension of the provided geometry does not match the job dimension")
Model_FEM = MeshNN_3D(Mesh_object, n_components = 3)
Model_FEM.Freeze_Mesh()
Model_FEM.UnFreeze_FEM()
if not config["solver"]["FrozenMesh"]:
Model_FEM.UnFreeze_Mesh()
Model_FEM.RefinementParameters( MaxGeneration = config["training"]["h_adapt_MaxGeneration"],
Jacobian_threshold = config["training"]["h_adapt_J_thrshld"])
Model_FEM.TrainingParameters( loss_decrease_c = config["training"]["loss_decrease_c"],
Max_epochs = config["training"]["n_epochs"],
learning_rate = config["training"]["learning_rate"])
Model_FEM.train()
Model_FEM = Training_2_3D_FEM(Model_FEM, config, Mat)
training_time = Model_FEM.training_recap["Duration_tot"]
torch.save(Model_FEM.state_dict(), 'Trained_models/Experiment_2/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item())) # To save a full coarse model
numpy.save('Trained_models/Experiment_2/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item()) +"_time.npy", training_time)
# Pplot.ExportFinalResult_VTK(Model_FEM,Mat,config["postprocess"]["Name_export"]+"_"+str(Model_FEM.state_dict()["h_max"].item())+"_final_2")
Experiment 3#
single-level training
r-adaptive mesh
optimizers:
l-bfgs
Adam + l-bfgs
# Experiment
element_sizes = [1, 0.5, 0.25]
config["training"]["multiscl_max_refinment"] = 1
config["solver"]["FrozenMesh"] = False
optimizers = ["lbfgs","adam"]
for o in optimizers:
config["training"]["optimizer"] = o
for e in element_sizes:
config["interpolation"]["MaxElemSize2D"] = e
if config["interpolation"]["dimension"] == 1:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["geometry"]["A"] # Section area of the 1D bar
)
elif config["interpolation"]["dimension"] in [2, 3]:
try:
Mat = pre.Material( flag_lame = False, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["material"]["nu"] # Poisson's ratio
)
except:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["lmbda"], # First Lame's coef
coef2 = config["material"]["mu"] # Second Lame's coef
)
MaxElemSize = pre.ElementSize(
dimension = config["interpolation"]["dimension"],
L = config["geometry"]["L"],
order = config["interpolation"]["order"],
np = config["interpolation"]["np"],
MaxElemSize2D = config["interpolation"]["MaxElemSize2D"]
)
Excluded = []
Mesh_object = pre.Mesh(
config["geometry"]["Name"], # Create the mesh object
MaxElemSize,
config["interpolation"]["order"],
config["interpolation"]["dimension"]
)
Mesh_object.AddBorders(config["Borders"]["Borders"])
Mesh_object.AddBCs( # Include Boundary physical domains infos (BCs+volume)
config["geometry"]["Volume_element"],
Excluded,
config["DirichletDictionryList"]
)
Mesh_object.MeshGeo() # Mesh the .geo file if .msh does not exist
Mesh_object.ReadMesh() # Parse the .msh file
Mesh_object.ExportMeshVtk()
if int(Mesh_object.dim) != int(Mesh_object.dimension):
raise ValueError("The dimension of the provided geometry does not match the job dimension")
Model_FEM = MeshNN_3D(Mesh_object, n_components = 3)
Model_FEM.Freeze_Mesh()
Model_FEM.UnFreeze_FEM()
if not config["solver"]["FrozenMesh"]:
Model_FEM.UnFreeze_Mesh()
Model_FEM.RefinementParameters( MaxGeneration = config["training"]["h_adapt_MaxGeneration"],
Jacobian_threshold = config["training"]["h_adapt_J_thrshld"])
Model_FEM.TrainingParameters( loss_decrease_c = config["training"]["loss_decrease_c"],
Max_epochs = config["training"]["n_epochs"],
learning_rate = config["training"]["learning_rate"])
Model_FEM.train()
Model_FEM = Training_2_3D_FEM(Model_FEM, config, Mat)
training_time = Model_FEM.training_recap["Duration_tot"]
torch.save(Model_FEM.state_dict(), 'Trained_models/Experiment_3/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item())) # To save a full coarse model
Pplot.ExportFinalResult_VTK(Model_FEM,Mat,config["postprocess"]["Name_export"]+"_"+str(Model_FEM.state_dict()["h_max"].item())+"_final_3")
numpy.save('Trained_models/Experiment_3/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item()) +"_time.npy", training_time)
Experiment 4#
r-adaptive mesh in the last level only
multi-level training
optimizers:
l-bfgs
Adam + l-bfgs
# Experiment
element_size = 1
max_refinment = 3
config["solver"]["FrozenMesh"] = True
config["solver"]["UnfreezeFinal"] = True
config["interpolation"]["MaxElemSize2D"] = element_size
optimizers = ["lbfgs","adam"]
for o in optimizers:
config["training"]["optimizer"] = o
for refin in range(1,max_refinment+1):
config["training"]["multiscl_max_refinment"] = refin
if config["interpolation"]["dimension"] == 1:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["geometry"]["A"] # Section area of the 1D bar
)
elif config["interpolation"]["dimension"] in [2, 3]:
try:
Mat = pre.Material( flag_lame = False, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["E"], # Young Modulus
coef2 = config["material"]["nu"] # Poisson's ratio
)
except:
Mat = pre.Material( flag_lame = True, # If True should input lmbda and mu instead of E and nu
coef1 = config["material"]["lmbda"], # First Lame's coef
coef2 = config["material"]["mu"] # Second Lame's coef
)
MaxElemSize = pre.ElementSize(
dimension = config["interpolation"]["dimension"],
L = config["geometry"]["L"],
order = config["interpolation"]["order"],
np = config["interpolation"]["np"],
MaxElemSize2D = config["interpolation"]["MaxElemSize2D"]
)
Excluded = []
Mesh_object = pre.Mesh(
config["geometry"]["Name"], # Create the mesh object
MaxElemSize,
config["interpolation"]["order"],
config["interpolation"]["dimension"]
)
Mesh_object.AddBorders(config["Borders"]["Borders"])
Mesh_object.AddBCs( # Include Boundary physical domains infos (BCs+volume)
config["geometry"]["Volume_element"],
Excluded,
config["DirichletDictionryList"]
)
Mesh_object.MeshGeo() # Mesh the .geo file if .msh does not exist
Mesh_object.ReadMesh() # Parse the .msh file
Mesh_object.ExportMeshVtk()
if int(Mesh_object.dim) != int(Mesh_object.dimension):
raise ValueError("The dimension of the provided geometry does not match the job dimension")
Model_FEM = MeshNN_3D(Mesh_object, n_components = 3)
Model_FEM.Freeze_Mesh()
Model_FEM.UnFreeze_FEM()
if config["solver"]["UnfreezeFinal"] and config["training"]["multiscl_max_refinment"]==1:
Model_FEM.UnFreeze_Mesh()
if not config["solver"]["FrozenMesh"]:
Model_FEM.UnFreeze_Mesh()
Model_FEM.RefinementParameters( MaxGeneration = config["training"]["h_adapt_MaxGeneration"],
Jacobian_threshold = config["training"]["h_adapt_J_thrshld"])
Model_FEM.TrainingParameters( loss_decrease_c = config["training"]["loss_decrease_c"],
Max_epochs = config["training"]["n_epochs"],
learning_rate = config["training"]["learning_rate"])
Model_FEM.train()
Model_FEM = Training_2_3D_FEM(Model_FEM, config, Mat)
training_time = Model_FEM.training_recap["Duration_tot"]
torch.save(Model_FEM.state_dict(), 'Trained_models/Experiment_4/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item())) # To save a full coarse model
Pplot.ExportFinalResult_VTK(Model_FEM,Mat,config["postprocess"]["Name_export"]+"_"+str(Model_FEM.state_dict()["h_max"].item())+"_final_4")
numpy.save('Trained_models/Experiment_4/'+o+'_3D_Cube_'+ str(Model_FEM.state_dict()["h_max"].item()) +"_time.npy", training_time)