-
3.2.2. Experimental Evaluation
-
-
To align with the findings in [moreau_shapes_2022], we initially consider a sphere as the solid at the center of the fluid domain. The primal and dual problems are solved using \(\mathbb{P}_2\) continuous finite elements for velocity and \(\mathbb{P}_1\) continuous finite elements for pressure. The expansion problem solution is also expressed with a \(\mathbb{P}_1\) continuous finite elements. It is important to keep in mind that our assumptions differ from those in the paper [moreau_shapes_2022]. In particular, we assume that the \(\Gamma_B\) edge is infinitely far from the \(\Gamma_S\) edge, which is not achievable using finite elements. Consequently, the obtained results may differ. The results obtained for the GD method and the NSGF method will be presented.
-
-
-
2D simulation for \(K_{11}\) case :
-
-
-
We consider the 2D \(K_{11}\) resistance problem. The initial fluid domain is a square with a spherical hole at its center. The parameter values chosen are described in the following table and with the following initial domain.
-
-
-
-
-
-
Figure 10. Initial geometry for the 2D case of Stokes.
-
-
-Table 14. Geometric and physics parameters of the initial domain for the 2D case of Stokes. The viscosity of the fluid is designed by \(\mu\). \(C\) represents the center of the sphere and the box, \(R\) the radius of the sphere and \(L\) the side length of the box.
-
-
-
-
-
-
-Symbol |
-Value (dimensionless) |
-
-
-\(\mu\) |
-\(1\) |
-
-
-\(C\) |
-\((0,0)\) |
-
-
-\(R\) |
-\(1\) |
-
-
-\(L\) |
-\(10\) |
-
-
-
-
-
For this specific test case, we employed an initial discretization of \(h=0.2\) around the spherical surface and \(h=1\) around the square. The optimization parameters we selected were \(l=20\), \(a=b=0.5\), and \(c=1e3\). In the NSGF method, remeshing plays a crucial role, particularly in the 2D case. To avoid convergence issues, we fixed the mesh at the edge of the solid, as the deformation was not significant. Consequently, the initial mesh was much more refined, with \(h=0.05\) at the solid boundary.
-
-
-Table 15. Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(K_{11}\) in 2D.
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(16.7729\) |
-\(14.2940\) |
-\(5.2101e-05\) |
-\(1.0154e-4\) |
-\(500\) |
-
-
-NSGF |
-\(16.9087\) |
-\(14.5917\) |
-\(1.1554e-3\) |
-\(3.3883e-4\) |
-\(500\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 11. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 2D \(K_{11}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
-
-
-
Figure 12. Initial domain (left) and final domain (right) for the GD method.
-
-
-
The shape obtained in 2D closely resembles a rugby ball, as described in [moreau_shapes_2022]. It effectively preserves the initial volume and achieves a lower cost function compared to the initial domain. The results obtained with the two methods exhibit striking similarity.
-
-
-
3D simulations for various resistance problem :
-
-
-
The initial fluid representation is a cube with a spherical hole at its center.
-
-
-Table 16. Geometric and physics parameters of the initial domain for the 3D case of Stokes. The viscosity of the fluid is designed by \(\mu\). \(C\) represents the center of the sphere and the box, \(R\) the radius of the sphere and \(L\) the side length of the box.
-
-
-
-
-
-
-Symbol |
-Value (dimensionless) |
-
-
-\(\mu\) |
-\(1\) |
-
-
-\(C\) |
-\((0,0,0)\) |
-
-
-\(R\) |
-\(1\) |
-
-
-\(L\) |
-\(10\) |
-
-
-
-
-
The parameter values chosen are described in the table below.
-
-
-
-
-
-
Figure 13. Illustration of the stokes problem. The \(\Gamma_B\) boundary is fixed. The \(\Gamma_S\) boundary is movable in order to optimize the shape.
-
-
-
In the following sections, we investigate the cases \(K_{11}\), \(K_{12}\), \(Q_{11}\), \(Q_{12}\), and \(C_{11}\) to compare our results with those presented in [moreau_shapes_2022]. For all simulations except the \(Q_{12}\) case, we use the optimization parameters \(t=0.03\), \(l=20\), \(a=b=0.5\), and \(c=1e3\). In the \(Q_{12}\) case, we employ \(t=5e-4\), \(l=20\), \(a=b=0.5\), and \(c=1e5\).
-
-
-
\(K_{11}\) resistance problem :
-
-
-
Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(K_{11}\) in 3D.
-
-
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(29.6246\) |
-\(27.2228\) |
-\(2.9460e-4\) |
-\(5.0356e-4\) |
-\(125\) |
-
-
-NSGF |
-\(29.6246\) |
-\(27.6108\) |
-\(1.0395e-2\) |
-\(6.6623e-2\) |
-\(33\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 14. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 3D \(K_{11}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
\(K_{11}\) simulation for the GD method.
-
-
-
-
-
-
We successfully achieve the optimal rugby ball shape while preserving volume and reducing the cost function. The NSGF method seems to be highly sensitive to remeshing, particularly at the ends, resulting in a shape reminiscent of a lemon. As a result, the simulation needs to be terminated before the mesh collapses at the ends.
-
-
-
\(K_{12}\) resistance problem :
-
-
-Table 17. Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(K_{12}\) in 3D.
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(-0.0188\) |
-\(-18.3755\) |
-\(0.2041\) |
-\(9.1405e-2\) |
-\(45\) |
-
-
-NSGF |
-\(-0.0188\) |
-\(-17.9411\) |
-\(8.5382e-3\) |
-\(2.5186e-3\) |
-\(850\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 15. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 3D \(K_{12}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
\(K_{12}\) simulation for the GD method.
-
-
-
-
-
-
As depicted in the previous figures, the simulations are terminated before the cost function reaches a local minimum due to mesh collapse. There are minimal differences in the final shape obtained between the two methods. Nevertheless, we manage to achieve a shape that closely resembles the one presented in the reference article, although it should be noted that the problem we are studying differs slightly as the boundary is at a finite distance from the rigid object.
-
-
-
\(Q_{11}\) resistance problem :
-
-
-Table 18. Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(Q_{11}\) in 3D.
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(24.1304\) |
-\(16.9092\) |
-\(5.4293e-3\) |
-\(5.0132e-3\) |
-\(500\) |
-
-
-NSGF |
-\(24.1304\) |
-\(16.9188\) |
-\(5.7716e-3\) |
-\(1.7381e-3\) |
-\(873\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 16. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 3D \(Q_{11}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
\(Q_{11}\) simulation for the GD method.
-
-
-
-
-
-
In this case, premature termination of the simulation is less evident, and convergence is observed with good volume retention. The shapes obtained are virtually identical between the two resolution methods. However, a notable observation is that the ends of the object approach the edges of the cube until they collide with them. This phenomenon is not observed when the cube boundaries are assumed to be at infinity, as depicted in [moreau_shapes_2022].
-
-
-
\(Q_{12}\) resistance problem :
-
-
-Table 19. Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(Q_{12}\) in 3D.
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(-9.0952e-3\) |
-\(-323.3069\) |
-\(3.9189e-2\) |
-\(3.7323e-2\) |
-\(980\) |
-
-
-NSGF |
-\(-9.0952e-3\) |
-\(-234.4491\) |
-\(1.5241e-2\) |
-\(6.3734e-3\) |
-\(1000\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 17. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 3D \(Q_{12}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
\(Q_{12}\) simulation for the GD method.
-
-
-
-
-
-
The boundary of the cube plays a significant role in the obtained results. Contrary to the expected dumbbell shape, we observe the ends of the solid flattening, leading to mesh collapse and an abrupt termination of the simulation. Additionally, these ends are positioned too closely to the cube boundary, resembling the observations made in the \(Q_{11}\) case. The shapes obtained using the two different methods exhibit a notable similarity.
-
-
-
\(C_{11}\) resistance problem :
-
-
-Table 20. Comparison of the cost function between the initial domain and the final domain, the volume error between the two domains, the \(H^1\)-norm of the displacement field at the end and the number of iterations for the resistance problem \(C_{11}\) in 3D.
-
-
-
-
-
-
-
-
-
-
-Method |
-\(J(\Omega_0)\) |
-\(J(\Omega_{n_{final}})\) |
-\(||\Omega_0|-|\Omega_{n_{final}}||\) |
-\(||\theta_{n_{final}}||_{H^1}\) |
-\(n_{final}\) |
-
-
-GD |
-\(2.7004e-3\) |
-\(-1.2593\) |
-\(4.2977e-2\) |
-\(4.4303e-2\) |
-\(74\) |
-
-
-NSGF |
-\(2.7004e-3\) |
-\(-1.7085\) |
-\(9.7921e-4\) |
-\(1.2600e-3\) |
-\(450\) |
-
-
-
-
-
-
-
-
-
-
-
-
-
Figure 18. Evolution of the cost function, the volume of the domain \(\Omega_n\) and the \(H^1\)-norm of \(\theta_n\) for the 3D \(C_{11}\) resistance problem with the GD method (top) and the NSGF method (bottom).
-
-
-
\(C_{11}\) simulation for the GD method.
-
-
-
-
-
-
Mesh collapse continues to be evident in this case, mainly due to sharp edges forming in specific regions. However, the dynamics align with the description provided in [moreau_shapes_2022], and volume conservation is satisfactory. Interestingly, we observe the emergence of four blades on the solid, each with varying prominence. It is worth noting that the difference between the two methods is apparent in figure below, where the blades are not necessarily in the same positions.
-
-
-
-
-
-
Figure 19. The black wireform corresponds to the result obtained using the GD method. The grey surface is the shape obtained using the NSGF method.
-
-