Table 3 Summary of normalized wall time magnitude and relative error of ROM.

From: Enhancing high-fidelity nonlinear solver with reduced order model

Example

1

2.1

2.2

2.3

3

4

remark

Normalized wall time

FOM

\(O(10^{0})\)

\(O(10^{0})\)

\(O(10^{0})\)

\(O(10^{0})\)

\(O(10^{0})\)

\(O(10^{0})\)

Default initialization

FOM with

ROM assists

\(O(10^{-1})\)

+

\(O(10^{-4})\)

\(O(10^{-1})\)

+

\(O(10^{-5})\)

\(O(10^{-1})\)

+

\(O(10^{-5})\)

\(O(10^{-1})\)

+

\(O(10^{-6})\)

\(O(10^{-1})\)

+

\(O(10^{-5})\)

\(O(10^{-1})\)

+

\(O(10^{-5})\)

\(\approx O(10^{-1})\)

ROM

\(O(10^{-4})\)

\(O(10^{-5})\)

\(O(10^{-5})\)

\(O(10^{-6})\)

\(O(10^{-5})\)

\(O(10^{-5})\)

cGAN-ROM for Ex. 1

BBT-ROM for Ex. 2, 3

Average relative error of ROM compared to FOM

0.13%

10.35%

6.17 %

17.17%

0.17%, 0.79%

2.33%

  1. ROM uses much less computational time (at least four orders of magnitude less) than the FOM nonlinear solver; hence, using ROM to assist FOM reduces a computational cost by one order of magnitude. Normalized wall time is calculated by normalizing each wall time used by the highest wall time (i.e., FOM with default initialization). For Example 3, there are two primary variables, \(p_h\) and \(s_h\); therefore, we report two relative error values for these two variables, respectively.