Skip to content

Latest commit

 

History

History
65 lines (46 loc) · 2.03 KB

TEP010_relativity.rst

File metadata and controls

65 lines (46 loc) · 2.03 KB

TEP010: Consistent Treatment of relativistic effects

Status

In Progress:

Responsible

@chvogl, @unoebauer, @ssim

Branches and Pull requests

Related issue: #448

Description

As many other Monte Carlo-based radiative transfer approaches, TARDIS takes a so-called mixed-frame approach to the problem. This means, that the packet propagation is carried out in the so-called lab frame (LF) since distances and time intervals are easily determined here. Radiation-matter interactions, however, are carried out in the so-called co-moving frame (CMF), which is a locally defined reference frame, which is advected with the fluid flow.

Such a mixed-frame approach requires frequent transformations of important packet properties (e.g. packet frequencies, energies, directions, etc.) between the two frames. Transformation laws exists (see for example Mihalas & Mihalas 1984), but only a very small subset of those is currently taken into account in TARDIS, namely:

  • first-order Doppler effect for packet energies and frequencies

The following effects are neglected or only partially accounted for

  • angle aberration
  • transformation of opacity
  • correct frame treatment of Monte Carlo estimators

The collaboration should carefully assess if the current treatment is sufficient to guarantee satisfactory accuracy in the low and mildly-relativistic regime and decide which relativistic effects should be included in TARDIS.

Implementation

@chvogl will lead the implementation of part or all of the relevant changes as part of this Type IIP project.

Backward compatibility

The proposed improvements will change the results of a TARDIS calculation. In most cases, these changes should be small since TARDIS is designed for the applications in which relatistic effects are small or at most only important up to first order in v/c.

Alternatives

Leave as is and accept potentially large errors in limiting cases.