• PDF
• Cite
• Share
Article Contents  Article Contents

# Multilinear POD-DEIM model reduction for 2D and 3D semilinear systems of differential equations

The author is a member of Indam-GNCS, which support is gratefully acknowledged

• We are interested in the numerical solution of coupled semilinear partial differential equations (PDEs) in two and three dimensions. Under certain assumptions on the domain, we take advantage of the Kronecker structure arising in standard space discretizations of the differential operators and illustrate how the resulting system of ordinary differential equations (ODEs) can be treated directly in matrix or tensor form. Moreover, in the framework of the proper orthogonal decomposition (POD) and the discrete empirical interpolation method (DEIM) we derive a two- and three-sided model order reduction strategy that is applied directly to the ODE system in matrix and tensor form respectively. We discuss how to integrate the reduced order model and, in particular, how to solve the tensor-valued linear system arising at each timestep of a semi-implicit time discretization scheme. We illustrate the efficiency of the proposed method through a comparison to existing techniques on classical benchmark problems such as the two- and three-dimensional Burgers equation.

Mathematics Subject Classification: Primary: 37M99, 15A21, 15A24, 15A69 65N06.

 Citation: • • Figure 1.  Example 1: Average relative error 28 (left) and online computational time (right) of the reduced order model and the full order model for different values of $\tau$

Figure 2.  Example 2: $u_1(x,y,0.5)$ discretized with $n = 200$. The exact solution (left), the $\mathsf{ho-pod-deim}$ approximation (middle), and the relative error mesh between the two (right)

Figure 3.  Example 2: $u_2(x,y,0.5)$ discretized with $n = 200$. The exact solution (left), the $\mathsf{ho-pod-deim}$ approximation (middle), and the relative error mesh between the two (right)

Figure 4.  Example 2: The average relative error through $n_{\mathfrak t} = 2n$ timesteps between the $\mathsf{ho-pod-deim}$ approximation $\widetilde{\bf U}_1(t)$ ($\widetilde{\bf U}_2(t)$) and the exact solution $u_1(x,y,t)$ ($u_2(x,y,t)$)

Figure 5.  Example 2: A comparison of the time required offline for basis construction (left) and online for integration (right) between $\mathsf{ho-pod-deim}$ and $\mathsf{pod-deim}$  for increasing $n$

Figure 6.  Example 3: A comparison of the offline time for increasing dimension $n$, between $\mathsf{ho-pod}$ and $\mathsf{pod}$ (left) and $\mathsf{ho-deim}$ and $\mathsf{deim}$ (right)

Figure 7.  Example 3: A comparison of the time to solve all linear systems of the form 25, for different values of $\tau$, between $\mathsf{t3-sylv}$ and Vec-lin. The $x-$axis displays the maximum dimension of the three vectorized equations for different values of $\tau$

Table 1.  Example 1. Dim. of ${\textsf{ho-pod}}$ and ${\textsf{ho-deim}}$ bases obtained for different $\tau$. The full order model has dimension $n = 1200$

 $\tau$ ${\bf U}_i$ left dim. $\mathsf{ho-pod}$ right dim. $\mathsf{ho-pod}$ left dim. $\mathsf{ho-deim}$ right dim. $\mathsf{ho-deim}$ $10^{-2}$ ${\mathbf{U}}_1$ 7 7 11 11 ${\mathbf{U}}_2$ 9 10 - - $10^{-4}$ ${\mathbf{U}}_1$ 18 20 23 23 ${\mathbf{U}}_2$ 19 20 - - $10^{-6}$ ${\mathbf{U}}_1$ 31 33 32 34 ${\mathbf{U}}_2$ 29 31 - - $10^{-8}$ ${\mathbf{U}}_1$ 43 46 44 47 ${\mathbf{U}}_2$ 37 40 - -

Table 2.  A breakdown of the $\mathsf{(ho)-pod}$ and $\mathsf{(ho)-deim}$ basis dimensions and the memory requirements for four different state space dimensions. Note that $\tau = 1/n^2$

 $n$ $\mathsf{algorithm}$ ${\bf U}_i$ $\mathsf{pod dim.}$ $\mathsf{deim dim.}$ ${\textsf{offline}}$$\mathsf{memory} \mathsf{online}$$\mathsf{memory}$ $60$ $\mathsf{ho-pod-deim}$ ${\bf U}_1$ 9/9 18/18 $98 n$ $54n$ ${\bf U}_2$ 9/9 18/18 $98n$ $54n$ $\mathsf{pod-deim}$  ${\bf U}_1$ 5 14 $400n^2$ $19n^2$ ${\bf U}_2$ 4 14 $400n^2$ $18n^2$ $200$ $\mathsf{ho-pod-deim}$ ${\bf U}_1$ 13/13 24/25 $153n$ $75n$ ${\bf U}_2$ 12/12 24/25 $153n$ $73n$ $\mathsf{pod-deim}$  ${\bf U}_1$ 9 23 $400n^2$ $32n^2$ ${\bf U}_2$ 8 23 $400n^2$ $31n^2$ $600$ $\mathsf{ho-pod-deim}$ ${\bf U}_1$ 16/17 32/32 $196n$ $97n$ ${\bf U}_2$ 16/16 32/32 $194n$ $96n$ $\mathsf{pod-deim}$  ${\bf U}_1$ 15 28 $400 n^2$ $43n^2$ ${\bf U}_2$ 14 28 $400n^2$ $42n^2$ $1200$ $\mathsf{ho-pod-deim}$ ${\bf U}_1$ 19/19 36/39 $219n$ $113n$ ${\bf U}_2$ 19/19 36/39 $215n$ $113 n$ $\mathsf{pod-deim}$  ${\bf U}_1$ 19 31 $400n^2$ $50 n^2$ ${\bf U}_2$ 18 31 $400 n^2$ $50 n^2$

Table 3.  Example 3. Dim. of $\mathsf{ho-pod}$ and $\mathsf{ho-deim}$ bases and the average error at 300 timesteps for increasing $r$. The full order model has dimension $n = 150$ and $\tau = 10^{-4}$

 $r$ $u$ $k_1$ $k_2$ $k_3$ $p_1$ $p_2$ $p_3$ $\mathsf{error} \bar{\mathcal{E}}(\pmb{\mathcal{ U}})$ $10$ $u_1$ 4 7 10 7 12 16 $1\cdot 10^{-4}$ $u_2$ 7 7 7 9 12 13 $6\cdot 10^{-5}$ $u_3$ 8 12 8 9 16 13 $1\cdot 10^{-4}$ $100$ $u_1$ 6 11 15 10 17 20 $3\cdot 10^{-5}$ $u_2$ 10 11 11 12 17 17 $4\cdot 10^{-5}$ $u_3$ 10 16 12 12 21 17 $4\cdot 10^{-5}$ $500$ $u_1$ 9 15 19 13 23 26 $2\cdot 10^{-5}$ $u_2$ 11 16 17 14 23 23 $3\cdot 10^{-5}$ $u_3$ 12 19 16 14 25 23 $4\cdot 10^{-5}$

Table 4.  Example 3. Memory and CPU time required for basis construction and integration. The full order model has dimension $n = 150$ and $\tau = 10^{-4}$

 r Online memory Basis time(s) FOM time(s) ROM time(s) 10 $177n$ 20 1641 1.9 100 $245n$ 20 1641 2.2 500 $318n$ 20 1641 3.3

Table 5.  Example 4. Dim. of ${\mathsf{ho-pod}}$ and ${\mathsf{ho-deim }}$ bases and further computational detalis for $\tau = 10^{-2}$ and $n = 150$

 $r_0$ ${\bf U}_i$ $\mathsf{pod dim.}$ ($k_1/k_2/k_3$) $\mathsf{deim dim.}$ ($p_1/p_2/p_3$) $\mathsf{online}$ $\mathsf{memory}$ $\mathsf{online}$ $\mathsf{time (s)}$ $\mathsf{error}$ $0.1$ ${\mathbf{U}}_1$ 2/2/2 5/5/5 $21n$ 1.29 $3\cdot10^{-4}$ ${\mathbf{U}}_2$ 2/2/2 3/3/3 $15n$ 1.20 $4\cdot10^{-4}$ ${\mathbf{U}}_3$ 18/18/18 - $54n$ 1.50 $3\cdot10^{-4}$ ${\mathbf{U}}_4$ 9/9/9 - $27n$ 0.63 $1\cdot10^{-2}$ $0.3$ ${\mathbf{U}}_1$ 8/8/8 9/9/9 $51n$ 1.56 $3\cdot10^{-4}$ ${\mathbf{U}}_2$ 8/8/8 9/9/9 $51n$ 1.13 $3\cdot10^{-4}$ ${\mathbf{U}}_3$ 43/43/42 - $128n$ 11.25 $3\cdot10^{-3}$ ${\mathbf{U}}_4$ 31/31/30 - $92n$ 4.27 $5\cdot10^{-3}$
• Figures(7)

Tables(5)

## Article Metrics  DownLoad:  Full-Size Img  PowerPoint