Table 5 Computation times using a GPU for the D-LSPF applied to the multi phase leak localization test case. All timings are computed with 5000 particles.

From: The deep latent space particle filter for real-time data assimilation with uncertainty quantification

 

HF

Reg-ViT-Trans

NoReg-ViT-Trans

Reg-Conv-Trans

Reg-ViT-NODE

P-RRMSE \(\downarrow\)

\(5.2\times 10^{-1}\)

\(\mathbf {9.6\times 10^{-3}}\)

\(1.1\times 10^{-2}\)

\(7.4\times 10^{-1}\)

\(1.5\times 10^{-2}\)

P-Wasserstein-1 \(\downarrow\)

169.9

168.4

875.7

172.7

S-RRMSE \(\downarrow\)

\(7.9\times 10^{-2}\)

\(\mathbf {2.5\times 10^{-2}}\)

\(2.9 \times 10^{-2}\)

\(8.3 \times 10^{-2}\)

\(3.1\times 10^{-2}\)

S-AMRMSE \(\downarrow\)

\(\mathbf {4.3\times 10^{-3}}\)

\(4.4\times 10^{-3}\)

\(4.5\times 10^{-3}\)

\(4.5\times 10^{-3}\)

P-NLL \(\downarrow\)

5.09

5.70

6.01

12.54

S-NLL \(\downarrow\)

16.87

15.37

18.37

17.32

Time (GPU) \(\downarrow\)

24.89 s

24.90 s

23.58 s

6.52

Speed-up (GPU) \(\downarrow\)

1807.95

1807.23

1908.40

6901.84

Time (CPU) \(\downarrow\)

45,000 s

332.3 s

317.0 s

328.8 s

45.9 s

Speed-up (CPU) \(\downarrow\)

135.4

142.0

136.9

980.8

  1. The high-fidelity solver makes use of 100 CPU cores and the neural network uses one GPU. “P-” refers to parameter estimation and “S-” refers to state estimation. RRMSE is the relative root mean squared error, the AMRMSE is the average moment RMSE with respect to the high-fidelity particle posterior, and NLL is the negative log-likelihood with respect to the high-fidelity particle posterior.
  2. Best values are in bold.