Deep material network (DMN) is a data-driven surrogate
model for heterogeneous materials (Liu and Wu,
2019). Based on linear elastic training data generated by high-fidelity
computational homogenization (such as FE-RVE
simulation), DMN is capable of predicting their elastic
and inelastic behaviors with high accuracy. DMN learns
the morphological characterizations of the microstructure by the homogenized effective
stiffness tensor and the stiffness of the constituents and . The effective stiffness tensor is usually computed using a
high-fidelity homogenization method such as the FE-RVE
method. The coefficients of DMN are computed by
minimizing the difference between the DMN and the
FE-RVE predictions, as shown in Figure 1.
Figure 1. Training of DMN.
The DMN model is available as built-in user-defined
material options. To activate the capability, you must specify a specific material user
type label in the material definition:
Use "DMN_UD" if the composite is reinforced with
unidirectional fibers.
Use "DMN_CHOPPED" if the composite is reinforced with
dispersed short fibers.
In this case, the dispersion of the fiber is defined
with a second-order orientation tensor.
Abaqus internally creates state variables used for this material model. You should not
define any solution-dependent state variables because they are reserved for specific
output variables.
A dedicated collection of parameter and property tables is available to include all
definitions required to use the model. You can use the abaqus
fetch utility to obtain the file containing the type definitions for
the parameter and property tables used by the DMN
model:
abaqus fetch job=ABQ_DMN_types.inp
Input File Usage
Use the following options to start the definition of a unidirectional
fiber-reinforced composite based on built-in user-defined material
options:
The DMN model in Abaqus is a rotation-free DMN based on the model
proposed in Gajek and Böhlke (2020). Compared to the original rotation-based
DMN proposed in Liu and Wu
(2019), the rotation-free DMN introduces
arbitrary lamination directions as shown in Figure 2. The rotation-free DMN
also has significantly improved performance compared to the original rotation-based
DMN because it avoids rotation of the homogenized stiffness matrices.
Figure 2. Architecture of a rotation-free DMN.
The rotation-free DMN uses building blocks governed
by analytical micromechanics solutions. The network is a binary tree structure, with
the bottom layer containing nodes. Each node is a two-phase laminate separated by an interface
with a normal vector . The strain and stress relationship between a parent node and its
two child nodes follows the rule of mixtures:
and
where and are the local strains of the two materials, and and are the local stresses of the two materials. and are the volume fractions of materials 1 and 2 (). The interface condition is defined as
where
corresponding to the Mandel notation of the stress tensor . The relationship between the local strain in each material and
the laminate strain is
where
For elastic materials, the strain-stress relationship of the laminate
can be written as:
where the effective elastic stiffness matrix is:
Because DMN is built based on building blocks with
analytical homogenization solutions as described above, it inherently satisfies
conservation laws and thermodynamic consistency. For the same reason, although the
network is trained solely based on linear elastic responses, its results can also be
extrapolated to produce accurate nonlinear predictions.
Composites used in industry applications usually have varying volume fraction or
aspect ratios from the manufacturing process. The
DMN model implemented in Abaqus is fully parametric based on the work by Li
(2024). The coefficients of the DMN are
trained using a series of FE-RVEs of varying
geometric parameters such as the fiber volume fraction as shown in Figure 3.
You do not need to specify the coefficients for the parametric
DMN. Abaqus internally stores the pretrained coefficients for all the supported
microstructures. When you specify the type of the microstructure of the composite
with the user type label, Abaqus automatically selects the coefficients and uses them for homogenization. For some
microstructures, Abaqus stores coefficients trained with different levels, and you can specify a
nondefault value for the number of levels to achieve better accuracy with additional
performance cost.
Figure 3. Parametric DMN trained with
FE-RVEs of various volume fractions.
Input File Usage
Use the following options to define the total level of the
DMN model:
The definition of the DMN model includes the
microstructure characteristics and the constituent properties. The microstructure
characteristics include the volume fraction, aspect ratio, and the dispersion
(described by the orientation tensor) of the fiber. The volume fraction and aspect
ratio cannot be dependent on temperature or field variables. You can define the
spatial variation of the fiber dispersion by specifying a distribution in the
orientation definition (Defining Fiber Dispersion for Multiscale or Short-Fiber Reinforced Linear Composite Materials).
You can specify the shape of the fiber in the DMN
model when the user type label is DMN_UD.
Input File Usage
Use the following options to define the fiber volume fraction of a
unidirectional fiber-reinforced composite:
You can define only one matrix and one fiber material in the
DMN model in Abaqus. Dedicated property table types are available to define constitutive laws
supported for DMN. You must specify labels for
these property tables to distinguish between matrix and fiber properties. For the
fiber material, only the following properties are supported:
Isotropic elasticity
Transversely isotropic elasticity
Isotropic thermal expansion
Transversely isotropic thermal expansion
For the matrix material, both linear and nonlinear properties are supported including:
Isotropic elasticity
Isotropic thermal expansion
Linear viscoelasticity defined with Prony series
Isotropic plasticity with one of the following hardening laws: tabular,
power law, or exponential linear law
Rate-dependent yield including: Power law, Chaboche law, and Prandtl
law
Linear Elasticity
The DMN model requires the definition of linear
elasticity in both the fiber and the matrix material. You can define linear
elasticity in the constituents with dedicated property table types. You can use
dedicated labels to indicate whether the elastic response is instantaneous or
long-term in the matrix material.
Input File Usage
Use the following option to define the isotropic elasticity in the fiber
material:
You can define linear viscoelasticity in the matrix material. The
DMN model supports only linear
viscoelasticity defined in the time domain with Prony series. You can define the
Prony series with dedicated parameter table types.
Input File Usage
Use the following option to define the Prony series in the matrix
material:
You can define isotropic hardening in the matrix material with dedicated property
table types. You can specify the hardening with one of the following options:
define the static yield stress, as a tabular function of the equivalent plastic
strain, ;
define the static yield stress using a power law, ; or
define the static yield stress using an exponential linear function, .
Input File Usage
Use the following option to define hardening using the tabular function in
the matrix material:
You can define rate-dependent yield (Rate-Dependent Yield)
in the matrix material with dedicated property table types. You can define the
rate-dependent yield stress as a function of the equivalent plastic strain rate, , and the static yield stress, , with one of the following options:
Example: Defining a Short Fiber-Reinforced Composite with
DMN
This example illustrates defining a short fiber-reinforced composite with
DMN.
MATERIAL, USER TYPE=DMN_CHOPPED, NAME=matNameUSER MATERIAL, CONSTANTS=0
PROPERTY TABLE, TYPE="ABQ_MULTSCALE_Volfrac"Data line to specify the volume fraction of the fiberPROPERTY TABLE, TYPE="ABQ_MULTSCALE_AspectRatio"Data line to specify the aspect ratio of the fiberPROPERTY TABLE, TYPE="ABQ_MULTSCALE_Oritens"Data line to specify the principal values of the fiber orientation tensorPROPERTY TABLE, TYPE="ABQ_MULTSCALE_IsoElas", LABEL="ABQ_MULTSCALE_Matrix_Elastic"Data line to define the linear elasticity of the matrix materialPROPERTY TABLE, TYPE="ABQ_MULTSCALE_ExpLinear", LABEL="ABQ_MULTSCALE_Matrix_Plastic"Data line to define the exponential linear hardening of the matrix materialPROPERTY TABLE, TYPE="ABQ_MULTSCALE_IsoElas", LABEL="ABQ_MULTSCALE_Fiber_Elastic"Data line to define the linear elasticity of the fiber material
The DMN material model is supported only in stress
analyses.
Output
For DMN material, Abaqus internally create SDVs for
specific output.
In addition to the standard output identifiers available in Abaqus (Abaqus/Standard Output Variable Identifiers), the
following variable has special meaning for the DMN
material model:
SDV1
Volume-averaged equivalent plastic strain, . User label
SDV_PEEQ_AVG.
You can also use a dedicated parameter table to request additional micro-level
output. When micro-level output is requested, the following variables have special
meaning for the DMN material model:
SDV1
11-component of stress in the matrix, .
SDV2
22-component of stress in the matrix, .
SDV3
33-component of stress in the matrix, .
SDV4
12-component of stress in the matrix, .
SDV5
13-component of stress in the matrix, .
SDV6
23-component of stress in the matrix, .
SDV7
11-component of stress in the fiber, .
SDV8
22-component of stress in the fiber, .
SDV9
33-component of stress in the fiber, .
SDV10
12-component of stress in the fiber, .
SDV11
13-component of stress in the fiber, .
SDV12
23-component of stress in the fiber, .
SDV13
Volume-averaged equivalent plastic strain, . User label
SDV_PEEQ_AVG.
Input File Usage
Use the following option to request micro-level output:
Gajek, S., M. Schneider, and T. Böhlke, “On the Micromechanics of Deep Material Networks,” Journal of the Mechanics and Physics of Solids, vol. 242, 2020.
Li, T., “Micromechanics-Informed Parametric Deep material network for physics behavior prediction of heterogeneous materials with a Varying Morphology,” Computer Methods in Applied Mechanics and Engineering, vol. 419, 2024.
Liu, Z., and C.T. Wu, “Exploring the 3D Architectures of Deep Material Network in Data-Driven Multiscale Mechanics,” Journal of the Mechanics and Physics of Solids, vol. 12720-46, 2019.
Liu, W., C.T. Wu, and M. Koishi, “A Deep Material Network for Multiscale Topology Learning and Accelerated Nonlinear Modeling of Heterogeneous Materials,” Computer Methods in Applied Mechanics and Engineering, vol. 3451138-1168, 2019.