search

LLNL / libROM

Model reduction library with an emphasis on large scale parallelism and linear subspace methods
https://www.librom.net
Other
197 stars 36 forks source link

Parametric dmdc heat conduction #264

Closed andersonw1 closed 3 months ago

andersonw1 commented 5 months ago

Adds parametric DMDc for heat conduction.

Example changing parameter in state vector: mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline --kappa 0.5 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline --kappa 0.75 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline --kappa 1.25 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline --kappa 1.5 -rdim 6

mpirun -np 8 parametric_dmdc_heat_conduction -s 1 --kappa 1 -online -predict -visit

Output: Relative error of DMDc temperature (u) at t_final: 0.5 is 0.0025944158 Elapsed time for solving FOM: 1.006058e+00 second Elapsed time for training DMDc: 7.535820e-01 second Elapsed time for predicting DMDc: 3.593333e-03 second

Example changing parameter in control vector: mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline -amp-in 1 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline -amp-in 2 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline -amp-in 4 -rdim 6 mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -offline -amp-in 5 -rdim 6

mpirun -np 8 parametric_dmdc_heat_conduction -s 1 -amp-in 3 -online -predict -visit

Output: Relative error of DMDc temperature (u) at t_final: 0.5 is 0.0028885833 Elapsed time for solving FOM: 9.007740e-01 second Elapsed time for training DMDc: 7.236116e-01 second Elapsed time for predicting DMDc: 5.489291e-03 second

Some things I am suspicious of: 1) Changing the amplitudes of the sources seems to have very little effect on the final accuracy of the ROM. For example, in the second example above, if we predict:

mpirun -np 8 parametric_dmdc_heat_conduction -s 1 --kappa 1 -amp-in 1000 -online -predict

the relative error for the final temperature is 0.013669751 despite the fact that -amp-in 1000 is far outside our trained values for the inlet amplitude.

2) I changed the if statements in matrixinterpolator.cpp. I think other examples that call the matrix interpolator should still work with the new logic, since we are now just checking if dimensions match. I still need to verify this.

dylan-copeland commented 4 months ago

There is a lot of file output that is hard to delete. It would be more convenient to use a subdirectory for file output, as done in examples/dmd/de_dg_advection_greedy.cpp (search for io_dir).

andersonw1 commented 4 months ago

There is a lot of file output that is hard to delete. It would be more convenient to use a subdirectory for file output, as done in examples/dmd/de_dg_advection_greedy.cpp (search for io_dir).

The new commit adds an output directory using the same "io_dir" code as in the other example.