Stress-constrained topology optimization using the constrained natural element method

Abstract

This paper presents a topology optimization framework to achieve volume minimization with crossland fatigue constraints under proportional loading using the constrained natural element method. The local minimization problem is solved by means of the augmented Lagrangian method to deal with a large number of evaluation points. To suppress the numerical instabilities, a neighbor-based filter is proposed and compared to the widely used density filter. Furthermore, to circumvent the problem of stress singularity, several different relaxations of the constraints are also investigated. Compared with the topology optimization procedure based on the finite element method, the proposed method has the advantage of showing greater flexibility and convenience in discretization of complex design domains and the ability to maintain stable output under various discretization conditions.

Appendices

Appendix 1: Sensitivity analysis of stress-constrained topology optimization

The sensitivity of the AL function in Eq. 19 at k-th iteration can be obtained as follows:

$$\begin{aligned} \frac{\partial J}{\partial \rho _{i}}=\sum _{j=1}^{N}\left( \frac{\partial f}{\partial \tilde{\rho }_{j}} \frac{\text{{d}}\tilde{\rho }_{j}}{\text{{d}}\rho _{i}}+\frac{1}{N} \frac{\partial P}{\partial \tilde{\rho }_{j}} \frac{\text{{d}} \tilde{\rho }_{j}}{\text{{d}}\rho _{i}}\right), \end{aligned}$$

(29)

where

$$\begin{aligned} P^{(k)}(\overline{\varvec{\rho }}, \varvec{u})=\sum _{j=1}^{N}\left[ \lambda _{j}^{(k)} h_{j}(\overline{\varvec{\rho }}, \varvec{u})+\frac{\mu ^{(k)}}{2} h_{j}(\overline{\varvec{\rho }}, \varvec{u})^{2}\right] \end{aligned}$$

(30)

is the penalization term.

For the first item on the right side of Eq. 29, $\text{{d}}\tilde{\rho }_{j}/\text{{d}}\rho _{i}=\mathcal {F}_{j i}$ is obtained from the filter relation in Eq. 10 and the term $\text{{d}}f /\text{{d}} \tilde{\rho }_{j}$ is given by

$$\begin{aligned} \frac{\partial f}{\partial \tilde{\rho }_{j}}=\frac{\partial }{\partial \tilde{\rho }_{j}} \frac{\sum _{i=1}^{N} \bar{\rho }_{i} V_{i}}{|\Omega |}=\frac{V_{j}}{|\Omega |} \frac{\text{{d}}\bar{\rho }_{j}}{\text{{d}} \tilde{\rho }_{j}}=\frac{V_{j}}{|\Omega |} \frac{\beta \left[ 1-\tanh \left( \beta \left( \tilde{\rho }_{j}-\eta \right) \right) \right] ^{2}}{\tanh (\beta \eta )+\tanh [\beta (1-\eta )]}. \end{aligned}$$

(31)

The calculation of the second term $\partial P/\partial \tilde{\rho }_{i}$ in Eq. 29 is expressed as follows:

$$\begin{aligned} \frac{\partial P^{(k)}}{\partial \tilde{\rho }_{j}}=\sum _{i=1}^{N}\left[ \lambda _{i}^{(k)}+\mu ^{(k)} h_{i}\right] \frac{\partial h_{i}}{\partial \tilde{\rho }_{j}}. \end{aligned}$$

(32)

Based on Eqs. 20 and 23e, the nonzero part of Eq. 32 is determined as follows:

$$\begin{aligned} \begin{aligned} \frac{\partial h_{i}}{\partial \tilde{\rho }_{j}}&= p(1-\epsilon ) \bar{\rho }_{i}^{p-1} \delta _{i j} g_{i}\left( g_{i}^{2}+1\right) \frac{\text{{d}}\bar{\rho }_{i}}{\text{{d}}\tilde{\rho }_{i}} \\&\quad +\frac{\epsilon +(1-\epsilon ) \bar{\rho }_{i}^{p}}{\sigma _{l i m}}\left( 3 g_{i}{ }^{2}+1\right) \left( \frac{\partial \sigma _{vm}^{i}}{\partial \varvec{u}}\right) ^{T} \frac{\partial \varvec{u}}{\partial \tilde{\rho }_{j}}, \end{aligned} \end{aligned}$$

(33)

when $\widetilde{g_{i}}(\overline{\varvec{\rho }}) \ge -\lambda _{i}^{(k)} / \mu ^{(k)}$, otherwise $\partial h_{i}/\partial \tilde{\rho }_{j}=0$.

The adjoint method is used here to reduce the cost of sensitivity evaluation, then $\partial P^{(k)} / \partial \tilde{\rho }_{j}$ is calculated as follows:

$$\begin{aligned} \frac{\partial P^{(k)}}{\partial \tilde{\rho }_{j}}=\left[ \lambda _{j}^{(k)}+\mu ^{(k)} h_{j}\right] p(1-\epsilon ) \bar{\rho }_{j}^{p-1} g_{j}\left( g_{j}^{2}+1\right) \frac{d \bar{\rho }_{j}}{\text{{d}}\tilde{\rho }_{j}}+\varvec{\eta }^{T} \frac{\partial \varvec{K}}{\partial \tilde{\rho }_{j}} \varvec{u} \end{aligned}$$

(34)

with the adjoint problem:

$$\begin{aligned} \varvec{\eta }^{T} \varvec{K}=-\sum _{i=1}^{N}\left[ \lambda _{i}^{(k)}+\mu ^{(k)} h_{i}\right] \frac{\epsilon +(1-\epsilon ) \bar{\rho }_{i}^{p}}{\sigma _{l i m}}\left( 3 g_{i}{ }^{2}+1\right) \left( \frac{\partial g_{i}}{\partial \varvec{u}}\right) \end{aligned}.$$

(35)

Since the stiffness matrix is constructed through Eq. 13, the second part on the right of Eq. 34 can be calculated as follows:

$$\begin{aligned} \varvec{\eta }^{T} \frac{\partial K}{\partial \tilde{\rho }_{j}} \varvec{u}=\varvec{\eta }_{j}^{T} \frac{\partial \varvec{K}_{j}}{\partial \tilde{\rho }_{j}} \varvec{u}_{j}=p(1-\epsilon ) \bar{\rho }_{j}^{p-1} \varvec{\eta }_{j}^{T} \varvec{K}_{j 0} \varvec{u}_{j} \frac{\text{{d}}\bar{\rho }_{j}}{\text{{d}}\tilde{\rho }_{j}} \end{aligned}.$$

(36)

The sensitivity of von Mises stress to the Cauchy stress vector is equal to

$$\begin{aligned} \frac{\partial \sigma _{vm}^{j}}{\partial \varvec{\sigma }}=\frac{\varvec{M} \varvec{\sigma }}{\sigma _{vm}^{j}}. \end{aligned}$$

(37)

The term $\partial \varvec{\sigma } / \partial \varvec{u}$ can be deduced through Eq. 6:

$$\begin{aligned} \frac{\partial \varvec{\sigma }}{\partial \varvec{u}}=\varvec{C} \varvec{B} \end{aligned}.$$

(38)

Finally, we obtain the following

$$\begin{aligned} \begin{aligned} \frac{\partial J}{\partial \rho _{i}}&= \left( \varvec{\mathcal {F}}^{T}\right) _{i}\left\{ \frac{V_{i}}{|\Omega |}+\frac{1}{N}\left[ \left( \lambda _{i}^{(k)}+\mu ^{(k)} h_{i}\right) p(1-\epsilon ) \bar{\rho }_{i}^{p-1} g_{i}\left( g_{i}^{2}+1\right) \right. \right. \\&\quad + \left. \left. p(1-\epsilon ) \bar{\rho }_{i}^{p-1} \varvec{\eta }_{i}^{T} \varvec{K}_{i 0} \varvec{u}_{i}\right] \right\} \frac{\beta \left[ 1-\tanh \left( \beta \left( \tilde{\rho }_{j}-\eta \right) \right) \right] ^{2}}{\tanh (\beta \eta )+\tanh [\beta (1-\eta )]} \end{aligned} \end{aligned},$$

(39)

where $\varvec{\mathcal {F}}_{i}$ is the ith row of the filter matrix.

Appendix 2: Sensitivity analysis of fatigue-constrained topology optimization

The AL-based method for topology optimization with fatigue constraints is similar to that with stress constraints, the main difference being that the load is cyclic in this case. Since the solid mechanics problem is linear elastic in this study, the relationships between the displacements and stresses obtained at each moment are also linear. Therefore, the equilibrium equation at any time point multiplied by an arbitrary adjoint vector $\varvec{\eta }$ is added to the penalization term when the adjoint method is used for sensitivity analysis:

$$\begin{aligned} \frac{\partial P^{(k)}}{\partial \tilde{\rho }_{j}}=\left[ \lambda _{j}^{(k)}+\mu ^{(k)} h_{j}\right] p(1-\epsilon ) \bar{\rho }_{j}^{p-1} g_{j}\left( g_{j}^{2}+1\right) \frac{\text{{d}}\bar{\rho }_{j}}{d \tilde{\rho }_{j}}+\varvec{\eta }^{T} \frac{\partial \varvec{K}}{\partial \tilde{\rho }_{j}} \varvec{u}_{r e f}, \end{aligned}$$

(40)

where $\varvec{u}_{r e f}$ represents the reference displacement vector at any predetermined loading moment. The adjoint problem then becomes:

$$\begin{aligned} \varvec{\eta }^{T} \varvec{K}=-\sum _{i=1}^{N}\left[ \lambda _{i}^{(k)}+\mu ^{(k)} h_{i}\right] \left( \frac{\partial h_{i}}{\partial \varvec{u}_{r e f}}\right) \end{aligned},$$

(41)

when $\widetilde{g_{i}}(\overline{\varvec{\rho }}) \ge -\lambda _{i}^{(k)} / \mu ^{(k)}$, $\partial h_{i} / \partial \varvec{u}_{r e f}$ is expressed as follows:

$$\begin{aligned} \frac{\partial h_{j}}{\partial \varvec{u}_{r e f}}=\frac{\partial \widetilde{g_{i}}}{\partial \sqrt{J_{2, a}^{j}}} \frac{\partial \sqrt{J_{2, a}^{j}}}{\partial \varvec{\sigma }_{a}} \frac{\partial \varvec{\sigma }_{a}}{\partial \varvec{u}_{r e f}}+\frac{\partial \widetilde{g_{i}}}{\partial \sigma _{H, \text{ max } }^{j}} \frac{\partial \sigma _{H, \text{ max } }^{j}}{\partial \sigma _{H, r e f}^{j}} \frac{\partial \sigma _{H, r e f}^{j}}{\partial \varvec{\sigma }_{r e f}} \frac{\partial \varvec{\sigma }_{r e f}}{\partial \varvec{u}_{r e f}} \end{aligned},$$

(42)

where $\varvec{\sigma }_{H,ref}$ and $\varvec{\sigma }_{ref}$ respectively represent the reference hydrostatic stress vector and the reference stress vector at any given reference time. Based on Eqs. 14 and 23e, the sensitivity of the constraint $\varvec{g}$ to the $\sqrt{J_{2, a}^{j}}$ and the $\varvec{\sigma }_{H,max}$ of Voronoi cell j are

$$\begin{aligned} \left\{ \begin{array}{c} \frac{\partial \widetilde{g_{i}}}{\partial \sqrt{J_{2, a}^{j}}} = \frac{p(1-\epsilon ) \bar{\rho }_{j}^{p-1}\left( 3 g_{j}^{2}+1\right) }{\beta _{c}} \\ \frac{\partial \widetilde{g_{i}}}{\partial \sigma _{H, \max }^{j}} = \frac{\alpha _{c} p(1-\epsilon ) \bar{\rho }_{j}^{p-1}\left( 3 g_{j}^{2}+1\right) }{\beta _{c}} \end{array}\right. \end{aligned}$$

(43)

The term $\partial \sqrt{J_{2, a}^{j}} / \partial \varvec{\sigma }_{a}$ in Eq. 43 is similar with Eq. 37, where $\partial \sqrt{J_{2, a}^{j}} / \partial \varvec{\sigma }_{a}=\varvec{M}_{c} \varvec{\sigma }_{a} / \sqrt{J_{2, a}^{j}}$.

The sensitivity of the stress amplitude to the reference displacement can be expressed as follows:

$$\begin{aligned} \begin{aligned} \frac{\partial \varvec{\sigma }_{a}}{\partial \varvec{u}_{r e f}}&=\frac{1}{2}\left( \frac{\partial \varvec{\sigma }_{max}}{\partial \varvec{u}_{r e f}}-\frac{\partial \varvec{\sigma }_{min}}{\partial \varvec{u}_{r e f}}\right) =\frac{1}{2}\left( \frac{\partial \varvec{\gamma }_{1} \varvec{\sigma }_{r e f}}{\partial \varvec{u}_{r e f}}-\frac{\partial \varvec{\gamma }_{2} \varvec{\sigma }_{r e f}}{\partial \varvec{u}_{r e f}}\right) \\&=\frac{1}{2}\left( \varvec{{\gamma }_{1}-\varvec{\gamma }_{2}}\right) \varvec{C} \varvec{B}, \end{aligned} \end{aligned}$$

(44)

where $\varvec{\gamma }_{1}=\varvec{\sigma }_{max} / \varvec{\sigma }_{ref}$ and $\varvec{\gamma }_{2}=\varvec{\sigma }_{min} / \varvec{\sigma }_{ref}$ are coefficient vectors.

For 2D plane stress state, $\partial \varvec{\sigma }_{H, r e f}^{j} / \partial \varvec{\sigma }_{r e f}^{j}=\left[ \begin{array}{lll} 1 / 3&1 / 3&0 \end{array}\right] ^{T}$.