exojax
exojax copied to clipboard
HajimeKawahara
🐈 Automatic differentiable spectrum modeling of exoplanets/brown dwarfs using JAX, compatible with NumPyro and JAXopt
I found that in some case, `trans_coeff` in the flux-adding reflection scheme becomes `nan`.
This issue requests ExoJAX's support for the Exomol broadener extension implemented in [radis #716](https://github.com/radis/radis/pull/716). Please refer to [radis #742](https://github.com/radis/radis/issues/742) as well.
For transmission spectroscopy, we found that there is essentially no method to significantly reduce device memory usage through layer-by-layer calculations. Therefore, a practical solution is to perform wavenumber-wise stitching. While...
I would figure out what parts limit the device memory use and computational costs of the current implementation for transmission spectra. - [x] a naive `opart` (layer-by-layer update of the...
Using `treeverse` for the computation of each layer might alleviate the memory-time tradeoff problem, so this will be investigated. It seems that `equinox` has implemented a `scan` with checkpointing functionality....
|R | 20000 | 40000 | 80000 | | ---- | ---- | ---- | ---- | |`opa`+`art` | | |  | | `opart` |  | | ...
The main methods in `atm` are currently `.run()`. But should we directly call the main method? Note that in any case we will keep `.run()` as an alias.
We would try to add the capability of an emulator-based opacity calculator, developing as [ExoEmulator](https://github.com/HajimeKawahara/exoemulator). Very experimental.
We do not yet have the rotation kernel for (high-res) transmission spectrum.
The `opa` class (in particular `OpaDirect`) is not yet compatible with atomic databases `adb`. This involves complex issues, so it is undecided whether it will be included in the `ExoJAX2`...
Metadata
Owner
Metadata
🐈 Automatic differentiable spectrum modeling of exoplanets/brown dwarfs using JAX, compatible with NumPyro and JAXopt