Nonlinear symmetry example
- Jobfile: nlsfile
- Skript: nonlin_symmetry
Porous medium equation
- Jobfile: pmefile
- Skript: newton
Plan
- make both runs work with the old scheme, somehow, adapt mu_ranges, nt, T and nelements accordingly
- try the same configuration with the combined basis generation approach using two sub-tests:
- with a factor of M=4*N,
- discarding the reduced basis vector if error grows. and without the combined basis generation approach
Testing
For better tests, I want to have a small sized problem for all the 6(4) problems, that run in a very short time.
Results
This is the results section.
Porous medium equation
- 's011n' : separate bases, normal size, adaptive basis generation, started with
newton([0,5],
false, 0, 1, 1,
'normal',
false, 4)
- EI epsilon of '1e-9' is reached after '140' basis functions 'stop_Mmax = 200', Lebesgue constant of '41.586' (max. poss.: '1.38e42')
- 3 parameter space adaptations after '20' basis extensions, '525' parameter samples at last refinement step.
- RB extension explodes after 26 extension steps, error estimator for 'N=26': '6e-2'
- minimum error estimator '9.28e-3' at step '14'.
- offline time: 6.9 h, four parallel computations
- 'c021nr': combined basis generation, two ei basis per extension steps, normal size, random parameter sampling (no adaptive parameter sampling)
newton([0,5],
true, 0, 2, 1,
'normal',
true, 4)
- EI basis of size '201', 'stop_Mmax = 200', last two skipped because of this, one skipped for too low residual.
- random parameter sample with '800' basis vectors
- RB extension stops with 'stop_Nmax = 50' at estimated error '8.329e-4'. (stop_epsilon = '1e-5')
- offline time: 9.9 h, four parallel computations
- 'c411n': combined basis generation, normal size, fixed 'M_by_N_ratio = 4', uniform parameter sampling (adaptive basis generation)
newton([0,5],
true, 4, 1, 1,
'normal',
false, 4)
- did not stop although 'stop_Nmax = 50' and 'stop_Mmax = 200' were reached. Re-run on pris with 'stop_Nmax = 120', 'stop_Mmax = 400'.
- Final result: Interrupted at 'N=58', 'M=230', because of increasing (exploding) error after 'epsilon = 5.3534e-04' has been reached at extension step '265' ('N=52', 'M=200') and just before last (fifth) parameter space refinement.
- Five parameter sample adaptations, finest level with '1968' elements.
- offline time: ? h, four parallel computations
- 'c021s': combined basis generation, small size, new algorithm for 'M_by_N_ratio = 0', uniform parameter sampling (adaptive basis generation)
newton([0,5],
true, 0, 2, 1,
'small',
false, 4)
- did not stop although 'stop_Nmax = 50' and 'stop_Mmax = 200' were reached. Re-run on lockheed with 'stop_Nmax = 120', 'stop_Mmax = 400'.
- Final result: Reached stop_epsilon barrier of '1e-5' with 'N=89', 'M=187'.
- Five parameter sample adaptations, finest level with '1905' elements.
- offline time: 8.4 h, four parallel computations
- 'c021n': combined basis generation, normal size, new algorithm for 'M_by_N_ratio = 0', uniform parameter sampling (adaptive basis generation)
newton([0,5],
true, 0, 2, 1,
'normal',
false, 4)
- did not stop although 'stop_Nmax = 50' and 'stop_Mmax = 200' were reached. Re-run on jax with 'stop_Nmax = 120', 'stop_Mmax = 400'. still running..., error explodes at 'N=73'?, okay the basis is discarded then, maybe it still works...
- Idea: Re-run with 'stop_epsilon = 1e-4' and/or lower 'ei_minimum_residual'
Nonlin-symmetry equation
- 's01ui': separate bases, implicit (Newton) discretization, started with
- stopped at 'stop_Nmax = 135'. Re-run on scorpion with 'stop_Nmax = 250'.
- Final results: epsilon limit 1e-5 reached with 'N=167', 'M=262',
- Lebesgue constant: '67.1493' of '7.4107e+78' maximum possible value.
- offline time: 2 h, four parallel processes
- No parameter adaptation!
- 'c02ui': combined basis generation, implicit (Newton) discretization, started with
- stopped at 'stop_Nmax = 135', 'stop_Mmax = 300', Re-run with 'stop_Nmax = 250', 'stop_Mmax = 400' on bigdaddy. Looks good...
- 'c31ui': combined basis generation, implicit (Newton) discretization, fixed 'M_by_N_ratio = 3',
- 'c02ue': combined basis generation, explicit discretization,