# Jacobian Definition

## Using Automatic Differentiation

One can elect to sacrifice some computing speed and calculate Jacobians automatically using automatic differentiation (AD). MOOSE employs the DualNumber class from the MetaPhysicL package in order to enable AD. If the application developer wants to make use of AD, they should inherit from ADKernel as opposed to Kernel. Additionally, when coupling in variables, the adCoupled* methods should be used. For example, to retrieve a coupled value, instead of using coupledValue("v") in the ADKernel constructor, adCoupledValue("v") should be used. adCoupledGradient should replace coupledGradient, etc. An example of coupling in an AD variable can be found in ADCoupledConvection.C and ADCoupledConvection.h. Moreover, material properties that may depend on the non-linear variables should be retrieved using getADMaterialProperty instead of getMaterialProperty. They should be declared in materials using declareADProperty. Example AD material source and header files can be found here and here; example kernel source and header files that use AD material properties can be found here and here.

Finite element shape functions are introduced in the documentation section shape functions. There we outline how our primary variables are summations of those shape functions multiplied by constant coefficients which are our degrees of freedom. In numerical implementation we give explicit illustration of how the derivative of a variable u with respect to its jth degree of freedom () is equal to the jth shape function . Similarly the derivative of with respect to is equal to . The code expression _phi[_j][_qp] represents in any MOOSE framework residual and Jacobian computing objects such as kernels and boundary conditions.

Any MOOSE kernel may have an arbitrary number of variables coupled into it. If these coupled variables use the same shape function family and order, then their associated s will be equivalent. However, if u and v use different shape functions then . As a developer, however, you do not in most cases have to worry about these differences in . MOOSE automatically updates the object member variable _phi to use the shape functions of the variable for whom the Jacobian is currently being computed. However, if the primary variable u is a scalar-valued (single-component) finite element variable and the coupled variable v is a vector-valued (multi-component) finite element variable (or visa versa), then you must introduce an additional member variable to represent the shape functions of the vector-valued (scalar-valued) variable. The name of this variable is up to the developer, but we suggest perhaps a _standard_ prefix for scalar valued finite-element variables and _vector_ for vector valued finite-element variables. The _standard_ prefix is suggested over _scalar_ so as not to be confused with a MooseVariableScalar, which only has a single value over the entire spatial domain. An example constructor for a standard kernel that couples in a vector-valued FE variable is shown below:


: Kernel(parameters),
_efield_id(coupled("efield")),
_efield(coupledVectorValue("efield")),
_efield_var(*getVectorVar("efield", 0)),
_vector_phi(_assembly.phi(_efield_var)),
_mobility(getParam<Real>("mobility"))
{
}


The associated declarations are:


const unsigned int _efield_id;
const VectorVariableValue & _efield;
VectorMooseVariable & _efield_var;
const VectorVariablePhiValue & _vector_phi;
const Real _mobility;
Real _sgn;


Residual, on-diagonal, and off-diagonal methods are respectively


Real
{
return -_grad_test[_i][_qp] * _sgn * _mobility * _efield[_qp] * _u[_qp];
}


and


Real
{
return -_grad_test[_i][_qp] * _sgn * _mobility * _efield[_qp] * _phi[_j][_qp];
}


and


Real
{
if (jvar == _efield_id)
return -_grad_test[_i][_qp] * _sgn * _mobility * _vector_phi[_j][_qp] * _u[_qp];
else
return 0;
}


An example constructor for a vector kernel that couples in a
scalar-valued FE variable is shown below:


const InputParameters & parameters)
: VectorKernel(parameters),
_v_id(coupled("v")),
_v_var(*getVar("v", 0)),
{
}


The associated declarations are:


const unsigned _v_id;
MooseVariable & _v_var;


Residual and off-diagonal Jacobian methods are respectively:


Real
{
}


and


Real

Note that only one member is needed to represent shape functions for standard MooseVariables and VectorMooseVariables. For example, if the vector-variables v and w are coupled into a standard kernel for u, only a single _vector_phi member needs to be added; there is not need for both a _v_phi and _w_phi. _vector_phi will be automatically updated to represent the shape functions for whichever vector variable the Jacobian is being computed for.