Python¶
Install¶
pip install llevant installs an abi3 wheel for Windows or Linux (the release page has
the same wheels); one wheel serves Python 3.9 and later, and numpy comes with it. For data
frames, add pip install polars or pandas. The wheel includes the GPU renderer and
the viewer:
python -m llevant render model.mcnp --gpu -o out.png # the `llevant` command
python -m llevant view model.mcnp # the viewer, from the wheel alone
python -m llevant view is the way to run the viewer where only Python packages can
be installed.
Other ways to install, and what your computer needs: Install. Building the module from source: contributing.
Use¶
tour.py
walks through the whole API. Run it as a script (python tour.py model.i out_dir), or open it
as a notebook: VS Code and PyCharm run its # %% cells, and
jupytext --to notebook tour.py makes an .ipynb.
import llevant, numpy as np
m = llevant.load("model.mcnp") # compiled once (cached on disk), then queried as often as needed
m.locate(0, 0, 0) # <Label inside 12 > 304[1,-2,0] > 7>: .cell, .material, .path...
r = m.classify(points) # (N, 3) array -> numpy columns: r.ids, r.cell, r.material, r.status
tr = m.trace_lines(starts, ends) # ragged: tr.offsets, tr.t0, tr.t1, tr.ids, tr.surface_in...
tr.lengths("material") # {material: total length}
m.slice(z=37, colour="material") # an Image: shown inline in Jupyter, .save("s.png"), .at(i, j)
m.render(direction=(1, 1, -0.5), hide=[12], gpu=True)
m.render(colour="realistic") # physically based; m.materials(), m.set_materials({12: "steel"})
m.volumes(h=2).by("material") # [Group(number, volume, sigma, bbox...)]
m.volumes_to(2, 1e-3, by="cell") # every cell, refined to sigma <= 0.1 % (Group.converged, .outcome)
m.check(h=2) # gaps and overlaps
m.diff(m.reload()) # what changed in the file, cell by cell
m.recompile(text) # a new version of the input, from text: only what changed is built
f = m.universe(20130) # a filler as a model of its own (cut="none" or a box)
r.to_polars(); tr.to_pandas() # data frames, with polars or pandas installed
m.locate(0, 0, 0).instance_path # "7 < 304[1 -2 0] < 12": which copy of the cell
f = llevant.Field.read(m, "run.msht", tally=14) # or Field.cells(m, {12: 3.5}), .instances, .mesh
m.slice(z=37, field=f.with_scale(1e-6, 1e-2)) # values over the cells
m.slice(z=37, field=f, outline_by="material") # lines between materials only
llevant.view("model.mcnp") # the viewer; returns when its window closes
llevant.view() # the viewer's start screen, to open a model from
- Numbers. Cells, materials, universes and surfaces are MCNP numbers.
- Inputs. Batch queries take anything
numpy.asarrayaccepts. C-ordered float64 is read in place; other layouts are copied by rows. - Outputs. Results are read-only numpy columns that Rust built: no Python object
per point or segment.
Labelobjects are made once per distinct label, on demand. - Threads. Every query runs in Rust, in parallel, without holding the GIL.
- The GPU model is uploaded on the first
render(gpu=True)and kept. - Material looks.
set_materialschanges how thatModelobject draws later images;slice(..., materials={12: "steel"})andrender(...)take looks for one image only. The compiled geometry is never changed. - Errors.
OSErrorfor files,ValueErrorfor bad arguments, andLlevantErrorfor other failures (the GPU).
Every function, method and property is described in the
Python API reference: units, MCNP numbers or indices,
and what None or -1 means. The same text ships with the module as type stubs
(_llevant.pyi), so editors show it as you type.