config.py

This file contains the hyper- and training parameters of the ICNN for learning the hidden material model using the NN-EUCLID framework. Table A.1 in the paper discusses these parameters.

Hyper- and training parameters

Parameter	Description
ensemble_size	Number of ICNNs in the ensemble
random_init	Randomly initialize weights and biases of ICNNs using xavier_uniform
n_input	Default = 3, i.e., the three principal invariants.
n_output	Default = 1, i.e., the strain energy density.
n_hidden	List of number of neurons for each hidden layer.
use_dropout	Use dropout in ICNN architecture.
dropout_rate	Dropout probability.
use_sftpSquared	Use squared softplus activation for the hidden layers.
scaling_sftpSq	Scale the output after squared softplus activation to mitigate exploding gradients.
opt_method	Specify the NN optimizer.
epochs	Number of epochs to train the ICNN
lr_schedule	Choose a learning rate scheduler to improve convergence and performance.
eqb_loss_factor	Factor to scale the force residuals at the free DoFs.
reaction_loss_factor	Factor to scale the force residuals at the fixed DoFs.
verbose_frequency	Prints the training progress every \(n^{th}\) (\(n=\)verbose_frequency) epoch.

Plotting parameters

Parameter	Description
plot_quantities	Which quantities to evaluate and plot.
strain_paths	Which strain paths to evaluate and plot.
lw_truth	Linewidth of the true strain energy density response.
lw_best	Linewidth of the strain energy response of accepted models.
lw_worst	Linewidth of the strain energy response of rejected models.
color_truth	Color for the true strain energy reponse
color_best	Color for the strain energy response of accepted models.
color_worst	Color for the strain energy response of rejected models.
alpha_best	Opacity of the line for the accepted models.
alpha_worst	Opacity of the line for the rejected models.
g_min, g_max	Range of the loading parameter gamma.
gamma_steps	How many loading steps to take.
remove_ensemble_outliers	If categorization into accepted and rejected should be made.
accept_ratio	Defines loss range in which accepted models fall into.
fs	fontsize