Activation functions used by feed-forward neural networks

This problem contains basic test cases for one or more Abaqus elements and features.

This page discusses:

Products Abaqus/Standard

Elements tested

T2D3

Features tested

Activation functions used by feed-forward neural networks using multiple neural networks defined in the same input file.

Problem description

Abaqus/Standard uses a simple two-dimensional wire model meshed with truss elements to test the activation functions supported by feed-forward neural networks.

The model contains multiple feed-forward neural networks. All neural networks are defined with both the input and output vectors having a dimension of one. Each neural network uses a single neuron with a different activation function. The neuron has a single weight with a unit value and the bias set to zero.

The analysis is performed over multiple static steps, with each step using a different neural network to prescribe the U2 displacements at all nodes in the model by evaluating the neural network from user subroutine DISP.

Results and discussion

Either the Fortran or the C/C++ version of user subroutine DISP can be used to apply the displacements during the analysis. Inside DISP, the step number is used as the ID of the neural network to be evaluated for calculating the displacement in the vertical direction at each node based on its x-coordinate in the initial configuration.

Figures

Figure 1. Deformed wire shapes at the end of each step corresponding to the activation function used during the step.