You can define feed-forward, dense (fully connected) neural networks in Abaqus/Standard and evaluate their output and derivatives using dedicated utility routines. You can
evaluate a neural network from any Abaqus/Standard user subroutine. Evaluating Abaqus Neural Networks describes the evaluation
utility routines. You can define multiple neural networks in a single input file. Each
neural network can have an arbitrary configuration.
Feed-Forward Neural Network Definition
To fully define a trained feed-forward neural network, you need to define the neural
network configuration and the neural network data (weights and biases). To define the
configuration of the network, you must specify the input and output size of the neural
network, the number of layers in the neural network, and the size and activation function
type for each layer. Some of the supported activation functions employ an activation
parameter with a default value. You can provide a nondefault value for this activation
parameter. You need to specify whether or not each layer uses biases.
To define the neural network data, you must provide the values of the weights and the
values of the biases (if used) for each layer. The neural network definition must explicity
specify all layers of the neural network. The size of the last layer ("output" layer) must
be identical to the neural network output size. Each neural network can have an arbitrary
number of layers. Abaqus/Standard assumes that the weights and biases are dimensionless.
You must assign a unique integer ID to each neural network
definition. This unique ID is used by the evaluation
utility routines to access the neural network to be evaluated. Optionally, you can assign a
(string) label to individual neural networks. However, you cannot use this label to access
the neural network during the evaluation.
Figure 1. Diagram of a feed-forward neural network with P layers, mapping an
n-dimensional input vector to an m-dimensional
output vector. Figure 2. Computational unit ("neuron") in a feed-forward neural network.
Activation Functions
The following activation functions are supported:
CELU (Continuously differentiable exponential linear
unit activation function) defined as .
You can specify a nondefault value for the activation parameter . The default value is .
ELU (Exponential linear unit activation function)
defined as .
You can specify a nondefault value for the activation parameter . The default value is .
FastTanh (Hyperbolic tangent approximation).
Gauss (Gauss activation function) defined as .
GELU (Gaussian error linear unit activation function)
defined using an approximation of .
LeakyReLU (Leaky rectified linear unit activation
function) defined as .
You can specify a nondefault value for the activation parameter . The default value is .
Linear (Linear activation function) defined as .
ReLU (Rectified linear unit activation function)
defined as .
Sigmoid (Logistic sigmoid activation function) defined
as .
SiLU (Sigmoid linear unit activation function) defined
as .
Sin (Sinusoid activation function) defined as .
SoftPlus (Softplus activation function) defined as .
SoftSign (Softsign activation function) defined as .
Swish (Swish activation function) defined as .
You can specify a nondefault value for the activation parameter . The default value is .
Tanh (Hyperbolic tangent activation function) defined
as .
All neurons in a layer use the same activation function. You can assign each layer a
different activation function.
Scaling
Abaqus/Standard assumes that the defined feed-forward neural networks are "trained." If any scaling of
the data sets was performed prior to training of the neural network, you must explicitly
account for the scaling effects (on the neural network inputs, outputs, and derivatives)
before and after the evaluation of the neural network in the user subroutines.
Usage
Input File Usage
Use the following options to define a feed-forward neural network.
A simple example of a feed-forward neural network using two layers (with two neurons
in the first layer and one neuron in the second layer, both layers using biases) and
approximating an XY curve:
A more complex example of a feed-forward neural network using three layers and mapping
an n-dimensional input quantity to an
m-dimensional output quantity (weights and biases are assumed to be
defined using eight values per data line):