matadi 0.0.3

Material Definition with Automatic Differentiation

matADi

Material Definition with Automatic Differentiation (AD)

matADi is a simple Python module which acts as a wrapper on top of casADi for easy definitions of hyperelastic strain energy functions. Gradients (stresses) and hessians (elasticity tensors) are carried out by casADi's powerful and fast Automatic Differentiation (AD) capabilities. It is designed to handle inputs with trailing axes which is especially useful for the application in Python-based finite element modules like scikit-fem or FElupe. Mixed-field formulations are supported as well as single-field formulations.

Installation

Install matADi from PyPI via pip.

pip install matadi

Usage

First, we have to define a symbolic variable on which our strain energy function will be based on.

Note: A variable of matADi is an instance of a symbolic variable of casADi (casadi.SX.sym). All matadi.math functions are simple links to (symbolic) casADi-functions.

from matadi import Variable, Material
from matadi.math import det, transpose, trace

F = Variable("F", 3, 3)

Next, take your favorite paper on hyperelasticity or be creative and define your own strain energy density function as a function of some variables x (where x is always a list of variables).

def neohooke(x, mu=1.0, bulk=200.0):
    """Strain energy density function of a nearly-incompressible 
    Neo-Hookean isotropic hyperelastic material formulation."""

    F = x[0]
    
    J = det(F)
    C = transpose(F) @ F
    I1_iso = J ** (-2 / 3) * trace(C)

    return mu * (I1_iso - 3) + bulk * (J - 1) ** 2 / 2

With this simple Python function we create an instance of a Material, which allows extra args and kwargs to be passed to our strain energy function. This instance now enables the evaluation of both gradient (stress) and hessian (elasticity) via methods based on automatic differentiation - optionally also on input data containing trailing axes. If necessary, the strain energy density function itself will be evaluated on input data with optional trailing axes by the function method.

Mat = Material(
    x=[F],
    fun=neohooke,
    kwargs={"mu": 1.0, "bulk": 10.0},
)

# init some random deformation gradients
defgrad = np.random.rand(3, 3, 5, 100) - 0.5

for a in range(3):
    defgrad[a, a] += 1.0

W = Mat.function([defgrad])[0]
P = Mat.gradient([defgrad])[0]
A = Mat.hessian([defgrad])[0]

Hints

Please have a look at casADi's documentation. It is very powerful but unfortunately does not support all the Python stuff you would expect. For example Python's default if-statements can't be used in combination with a symbolic boolean operation. If you use eigvals to symbolically calculate eigenvalues and their corresponding gradients please call gradient and hessian methods with Mat.gradient([defgrad], modify=True, eps=1e-5) to avoid the gradient to be filled with NaN's. This is because the gradient of the implemented eigenvalue calculation is not defined for the case of repeated equal eigenvalues. The modify argument adds a small number eps=1e-5 to the diagonal entries of the input data.