Decisions
Collision rates
- for MVP Eiron: use flat rate
- for general case: samples with spline interpolation is a general representation, as long as there is a possibility to use a log scale for some parts of the energy spectrum
- the runtime collision rate function can then use a single representation for the collision rate(energy) function, a preprocessing step transforms any specific fits to these samples to use in the spline interp
Trajectory representations
- for MVP Eiron: use point paths
- for EIRENE, a point path + sparse grid representation would be more efficient in certain cases, where the trajectory discretization requires a lot of checks that slow the process down, which makes storing the result worth it, despite the storage cost
- related: there are multiple neighbor-indexing solutions, EIRENE uses a neighbor list (graph array repr.), but an octree/quadtree solution could compete, is an something to think about in future projects
Grid representation
- for MVP Eiron: use cartesian 2 D grid
- for general case: the grid needs to be aligned with the magnetic field, which means that a polygonal mesh (or curved surfaces) are needed
- a polygonal mesh will be a general grid description that should work in all cases
- the walls can be a part of the polygonal mesh
- Two most used grid representations: LEVGEO=4 and LEVGEO=10, LEVGEO 4 is triangles, polygonal mesh, LEVGEO 10 is user provided geometric model, used by EMC3
(LEVGEO 5 is tetrahedrons, which would be the 3D case)
Estimates
- Particle density, momentum density, and event density is enough to calculate all other estimates
- Mean + variance is also enough, although a range (max,min) could be useful
- note: it should not be too difficult to add additional statistics accumulators to the estimates, so this should be a non-issue
Other comments
Domain decomposition (Dmitriy, Yannick):
- Dmitriy: Divertor slab in JET is a "good basis" for domain decomposition (?)
- translandtion: we should care about communication load when deciding the partitioning scheme, and a physics-informed partitioning (based on where particle flows are largely self-contained) is one starting point for domain-decomposition
Extra: Idea for handling curved surfaces (purely academic, not a serious suggestion for work)
- pretend it's straight, adjust the track length attributed in the cell by adding the error to the curved surface
- to make this accurate, you need interpolated smooth collision rates
- CON: that makes the integral more complicated
