Newton-Raphson method fails!

Question

I am trying to solve an equation like $R(x) = 0$, using Newton-Raphson method. To obtain the $x$ increment in each iteration I solve $dx = -(A)^{-1}\cdot R$ where $A = dR/dx$. But the convergence criteria, here $||R||$ reduces to a certain amount, depending on the input, e.g. 1.e-8, and it remains so. After much trying I did not find any special bug, so i am starting to have second thoughts about my implementation.

I am wondering now that the approach to calculate the inverse of matrix $A$ might cause inaccuracy so that the convergence fails. What do you think?

EDIT: The code is written in fortran; I calculate the inverse using MKL's routines: dgetrf and dgetri. I am using double-precision variables.

The problem is to solve $R(\sigma)$ in n steps using Backward-Euler method. So functions $Z^p$ and $Z^d$ are also dependent on $\sigma$. In each step $\Delta\epsilon_{n+1}$ is updated and we solve an implicit problem to update $\sigma$

where for a plane strain problem the stress tensor reads

$$ {\boldsymbol\sigma} = \left[ \begin{array}{c} \sigma_{11}\\ \sigma_{22}\\ \sigma_{33}\\ \sigma_{12} \end{array} \right] $$

The function $R$ is: $$ \mathbf{R}_{\boldsymbol\sigma} = {\boldsymbol\sigma}_{n+1} - {\boldsymbol\sigma}_n - \mathbf{D}\Delta{\boldsymbol\epsilon_{n}} + \Delta t\,\mathbf{D}\left[\mathbf{Z}^p({\boldsymbol\sigma}^{\rm{dev}}) + \mathbf{Z}^d({\boldsymbol\sigma}^{\rm{dev}})\right] $$ where $$ \begin{align} \left[\mathbf{Z}^d({\boldsymbol\sigma}^{\rm{dev}})\right]_i &= \alpha\,\left[{\boldsymbol\sigma}^{\rm{dev}}\right]_i \\ \left[\mathbf{Z}^p({\boldsymbol\sigma}^{\rm{dev}})\right]_i &= \beta\,||{\boldsymbol\sigma}^{\rm{dev}}||^{\frac{1-m}{m}} \left[{\boldsymbol\sigma}^{\rm{dev}}\right]_i \\ \left[{\boldsymbol\sigma}^{\rm{dev}}\right]_i &= \left[\mathbf{I}^{\rm{dev}}\right]_{ij}\,\left[{\boldsymbol\sigma}\right]_j \end{align} $$

$I^{dev}$ is defined as:

$$ I_{ij}^{dev} = \delta_{ij} - \frac{1}{3}m_im_j \quad\text{where}\quad \mathbf{m} = \left[ \begin{array}{c} 1\\ 1\\ 1\\ 0 \end{array} \right] $$

and $\delta$ is the Kronecker delta.

The tangent modulus is given by $$ d\mathbf{R}_{\boldsymbol{\sigma}} = \frac{\partial \mathbf{R}_{\boldsymbol{\sigma}}}{\partial {\boldsymbol{\sigma}}}\,d{\boldsymbol{\sigma}} = \left[\mathbf{I} + \Delta t\,\mathbf{D}\left(\frac{\partial \mathbf{Z}^d({\boldsymbol\sigma}^{\rm{dev}})}{\partial \boldsymbol{\sigma}} + \frac{\partial \mathbf{Z}^p({\boldsymbol\sigma}^{\rm{dev}})}{\partial \boldsymbol{\sigma}}\right) \right]\,d{\boldsymbol{\sigma}} = \mathbf{A}\,d{\boldsymbol{\sigma}} $$ If we expand $\mathbf{A}$ we get $$ \mathbf{A} = \mathbf{I} + \Delta t\,\mathbf{D}\left[\left(\alpha+\beta\,||{\boldsymbol\sigma}^{\rm{dev}}||^{\frac{1-m}{m}}\right)\,\mathbf{I}^{\rm{dev}} + \frac{(1-m)\beta}{m}\,||{\boldsymbol\sigma}^{\rm{dev}}||^{\frac{1-3m}{m}}{\boldsymbol\sigma}^{\rm{dev}}\otimes{\boldsymbol\sigma}^{\rm{dev}}\right] $$ Here $x=\sigma_{n+1}$. Also, $\alpha, \beta, \Delta t, I_{ij}^{dev}, D$ and $m$ are constant($m=0.3$ and $i,j=1,...,4$.). $\sigma_n$ and $\Delta{\boldsymbol\epsilon_{n}}$ are constant in step n.

Here you can see the results for $R$ and $\sigma_{n+1}$.

Answer 1

I suggest you do the following to check your implementation:

Take your nearly-converged sigma vector and then calculate the 4x4 Jacobian matrix numerically, column-by-column using a central difference approximation. You may have to experiment a bit with the step size to obtain an accurate result. This numerical Jacobian should be very close to your analytical one.

Answer 2

As a check we can look at the problem in terms tensors. Let $$ \boldsymbol{\sigma}^{\text{dev}} = \boldsymbol{\sigma} - \frac{1}{3}(\boldsymbol{\sigma}:\mathbf{I})\,\mathbf{I} $$ where $\mathbf{I}$ is the second-order identity tensor. Then the fourth-order tensor $\mathsf{I}^{\rm dev}$ is given by $$ \mathsf{I}^{\rm dev} = \mathsf{I}^{(4s)} - \frac{1}{3} \mathbf{I}\otimes\mathbf{I} $$ where $\mathsf{I}^{(4s)}$ is the symmetric fourth-order identity tensor with components $1/2(\delta_{ik}\delta_{jl} + \delta_{il}\delta_{jk})$ and $\delta_{ij}$ is the Kronecker delta.

We can also express $\mathbf{Z}^p$ in tensor form as $$ \boldsymbol{Z}^p = \beta\,\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-m}{m}} \boldsymbol{\sigma}^{\rm dev} \,. $$ Then $$ \frac{\partial \boldsymbol{Z}^p}{\partial\boldsymbol{\sigma}} = \beta\left[\frac{\partial}{\partial \boldsymbol{\sigma}}\left(\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-m}{m}}\right)\otimes\boldsymbol{\sigma}^{\rm dev} + \left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-m}{m}} \frac{\partial \boldsymbol{\sigma}^{\rm dev}}{\partial\boldsymbol{\sigma}}\right] $$ Let us look at the first term $$ \frac{\partial}{\partial \boldsymbol{\sigma}}\left(\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-m}{m}}\right) = \frac{1-m}{m}\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-2m}{m}} \frac{\partial\left\|\boldsymbol{\sigma}^{\rm dev}\right\|}{\partial \boldsymbol{\sigma}} $$ Next look at the part $$ \frac{\partial\left\|\boldsymbol{\sigma}^{\rm dev}\right\|}{\partial \boldsymbol{\sigma}} = \frac{1}{2\left\|\boldsymbol{\sigma}^{\rm dev}\right\|}\,\frac{\partial}{\partial\boldsymbol{\sigma}}(\boldsymbol{\sigma}^{\rm dev}:\boldsymbol{\sigma}^{\rm dev}) $$ Finally, we look at the last bit above, using $\boldsymbol{\sigma}^{\rm dev}:\mathbf{I} = 0$, $$ \frac{\partial}{\partial\boldsymbol{\sigma}}(\boldsymbol{\sigma}^{\rm dev}:\boldsymbol{\sigma}^{\rm dev}) = \mathsf{I}^{\rm dev}:\boldsymbol{\sigma}^{\rm dev} + \boldsymbol{\sigma}^{\rm dev}:\mathsf{I}^{\rm dev} = 2\boldsymbol{\sigma}^{\rm dev} $$ Therefore, $$ \frac{\partial \boldsymbol{Z}^p}{\partial\boldsymbol{\sigma}} = \frac{\beta(1-m)}{m}\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-3m}{m}} \boldsymbol{\sigma}^{\rm dev}\otimes\boldsymbol{\sigma}^{\rm dev} + \beta\,\left\|\boldsymbol{\sigma}^{\rm dev}\right\|^{\frac{1-m}{m}}\mathsf{I}^{\rm dev} \,. $$ This is different from the expression you have used. So you will need to recheck your algebra. So the tangent modulus appears to be correct.