Modeling Soft Matter Using Constitutive Artificial Neural Networks
(CANNs)
A special-purpose built-in material is available for modeling soft matter using
constitutive artificial neural networks (CANNs). You can use
this material model to simulate the response of biological tissue and other soft material systems
that can be described with an anisotropic hyperelastic strain energy potential.
Abaqus provides a general framework to integrate trained constitutive artificial neural networks
(CANNs) models into an Abaqus simulation with minimal user effort. CANNs are a general
class of anisotropic hyperelastic constitutive models that are described by a neural network
architecture. The architecture used in the AbaqusCANNs framework is based on the work of Linka et al. (2023)
and Peirlinck et al.
(2025).
To activate the framework, the material definition must include the
CANNs material user type. It must also include parameter tables
with a declared type that starts with "ABQ_CANN." In addition, you must
specify a user-defined anisotropic hyperelastic material definition, with the number of fiber
families, , based on the specific model at hand.
A dedicated collection of parameter tables is available to include all the definitions required
to model CANNs models. You can use the abaqus
fetch utility to obtain the file containing the type definitions of the parameter
tables used by the CANNs framework as
follows:
abaqus fetch job=ABQ_CANN_types.inp
Input File Usage
Use the following options to activate the compressible
CANNs framework in Abaqus/Standard:
The models available in the CANNs framework are described by
the neural network architecture shown in Figure 1. This feed-forward
architecture hardwires physical constraints into the network design using deformation invariants
as inputs and specific activation functions in the hidden layers. The output of the architecture
is the desired strain energy potential. Compared to classical neural network architectures,
CANNs are typically not fully connected by design to satisfy
the condition of polyconvexity before you run the analysis (see Linka et al.
(2023)).
Figure 1. Architecture of a constitutive artificial neural networks model.
CANNs Model Definition
The subsections below detail the invariant inputs and activation functions used to define the
CANNs material models.
Strain Energy Potential Output
The model defines a strain energy function, , as the sum of individual polyconvex functions of the inputs
where each contribution takes the form
In the above equation, is an invariant input; the integers , , are indices specifying the activation functions , , in layers 0, 1, and 2, respectively; and , , are the weights associated with these functions for input .
Invariant Inputs
The invariant inputs are “shifted” invariants defined by
where are isotropic () and anisotropic () invariants associated with the state of deformation of the material.
The number of anisotropic invariants depends on the total number of material fiber
directions; up to fiber directions can be considered. The invariants follow the enumeration
convention in Abaqus. The quantity represents the value of the corresponding invariant in the reference configuration. It follows that in the undeformed state. Table 1 lists the invariant inputs.
are first and second deviatoric strain invariants,
is the elastic volume ratio (or third strain invariant),
and
are pseudo-invariants of the distortional part of the right
Cauchy-Green strain , and
are geometric constants (independent of deformation) equal to the cosine of the angle
between the directions of any two families of fibers in the reference configuration.
The activation functions in the three hidden layers of this
CANNs architecture are the functions , , and , in layers 0, 1, and 2, respectively (see Figure 1). For each input, (and, therefore, index ), the integers , , are indices specifying these activation functions, and , , are the weights associated with these functions. That is, is the weight of layer 0 in layer 1, is the weight of layer 1 in layer 2, and is the weight of layer 2 in the definition of .
Table 2 defines the activation
functions as follows:
Layer 0 contains the identity , Macaulay bracket , and absolute value operators. The Macaulay bracket is defined as .
Layer 1 contains integer exponentiation functions of powers .
Layer 2 contains linear scaling, exponential and logarithmic functions. The definitions of
activation functions and include a scaling factor, , that makes them different from those in the work of Linka et al.
(2023).
Table 2. Activation functions.
Index , ,
Layer 0,
Layer 1,
Layer 2,
Strain Energy Potential Definition
It follows from the above definitions that each contribution to the CANNs strain energy potential is uniquely defined by the tuple
where
is the index specifying the input ,
is the index specifying the activation function in layer 0,
is the index specifying the activation function in layer 1,
is the index specifying the activation function in layer 2,
is the weight of layer 0 in layer 1,
is the weight of layer 1 in layer 2, and
is the weight of layer 2 in the definition of .
Abaqus/Explicit requires the definition of some compressibility. Therefore, the parameter table of type
ABQ_CANN_Coeff must include a contribution from the third invariant.
Input File Usage
Use the following options to define the strain energy potential, :
Repeat the data line as often as required to define the list of all the contributions .
Derived Invariant Inputs
Some forms of strain energy potentials, such as the Holzapfel-Gasser-Ogden form, are
expressed in terms of linear combinations of the 15 invariants in Table 1. Those combinations are referred to as
“derived invariants” and provide a level of interconnectivity between the 15 invariants in
layer 0 of the CANNs model. Derived invariant is defined by
and can be used as an input in the CANNs
architecture using the invariant input (see Table 3).
Table 3. Input enumeration for derived invariants.
Index
101
102
103
104
Input File Usage
Use the following options to define the derived invariants :
Repeat the data line as often as required to define the list of all the derived
invariants .
Examples
The examples below illustrate how to use CANNs models for the
case of well-known forms of hyperelastic strain energy potentials.
Neo-Hooke Form
The Neo-Hooke strain energy potential is given by
which can be rewritten as
in terms of the CANNs input invariants using Table 1.
Each of these contributions can be expressed in its tuple definition :
The contribution of the first invariant () can be expressed as . This corresponds to using the linear function in each layer () with weights and set to 1.0 and weight set to .
The contribution of the third invariant () can be expressed as . This corresponds to using the linear function in layers 0 and 2 () and the quadratic function in layer 1 (). Weights and set to 1.0 and weight set to .
Input File Usage
Use the following options to define the Neo-Hooke strain energy potential as a
CANNs model:
Holzapfel-Gasser-Ogden Form (with Fiber Dispersion )
The Holzapfel-Gasser-Ogden strain energy potential with fiber dispersion, , is given by
with
This can be rewritten as
in terms of the CANNs input invariants using Table 1 and of and , two derived invariant inputs defined by
representing the effect of fiber dispersion.
Input File Usage
Use the following options to define the Holzapfel-Gasser-Ogden strain energy potential
with fiber dispersion, , as a CANNs model with derived
invariants:
You can use the CANNs framework with all plane stress, plane
strain, and 3D stress-displacement elements in both Abaqus/Standard and Abaqus/Explicit. This set of eligible elements includes 3D solids, 2D axisymmetric, and generalized plane
strain elements.
In Abaqus/Standard, you can use hybrid elements to model a fully incompressible
CANNs material model. You cannot use compressible anisotropic
hyperplastic material behavior (including compressible CANNs
models) with hybrid elements. For reduced-integration elements, you must specify the hourglass
stiffness because Abaqus does not calculate the default values for CANNs models.
In Abaqus/Standard and Abaqus/Explicit, the transverse shear stiffness definition is required for shell elements since Abaqus does not calculate the default values for CANNs models.
References
Linka, K., and E. Kuhl, “A New Family of Constitutive Artificial Neural Networks Toward Automated Model Discovery,” Computer Methods in Applied Mechanics and Engineering, vol. 403115731, 2023.
Peirlinck, M., J. A. Hurtado, M. K. Rausch, A. Buganza Tepole, and E. Kuhl, “A Universal Material Model Subroutine for Soft Matter Systems,” Engineering with Computers, vol. 41, pp. 905–927, 2025.