exojax
exojax copied to clipboard
HajimeKawahara
Premodit Elower grid error rethinking
Previously, I estimated a maximum error of delta E_lower. But, it looks too pessimistic. To reduce more memory use, we need to rethink it.
Analytic estimate of the error of line strength using PreMODIT Elower grid:
no dependence on Elower_line
I find that I have not checked if a simple T gird instead of Elower in PreMODIT works or not. We'll see.
Errors in LSD as a function of interval_contrast
Tguess = 1000K, comparison with MODIT LSD
| T (K) | ic=0.1 | ic=0.2 | ic=0.3 |
|---|---|---|---|
| 1200K | 0.3% | 1% | 3% |
| 700K | 0.5% | 2% | 5% |
| T (K) | ic=0.1 | ic=0.2 | ic=0.3 |
|---|---|---|---|
| 1200K |  |
 |
 |
| 700K |  |
 |
 |
The above results are completely consistent with the analytic estimate (elower_grid.ipynb)
- ic=0.2 case
short summary
interval_contrast
- the errors from the comparison with MODIT (H2O/POKAZATEL) can be reproduced by the analytical model. So, the reason is from the deviation of the (weighted) linear interpolation of two Elower grid points.
-
interval_constant = 0.1keeps the LSD error within 1% for 250-2000K, assuming Tgue=1000K
elower_grid.ipynb
some indication
We need ~200 grids for Elower in the case of H2O/POKAZATEL, which has large Elower in databases.
See the Elower distribution.
freq_elower.ipynb
I feel NE=200 is a bit large. Maybe #317 relaxes the number of grids.. I hope.
I think I did improve the memory use of the E lower grid (Note that we need to multiply 2 to NE to compare the memory use because the derivative of the weight is used)
exojax/tests/integration/comparison/premodit/fig_elower_grid_error.py