Gmsh#

.msh read + write (v2) eager

Summary of the specification#

A Gmsh .msh file is a sequence of $Section / $EndSection blocks. $MeshFormat states the version, whether the file is ASCII or binary, and the size of a float. $Nodes lists node tags with coordinates, and $Elements lists elements with an integer type code and their node tags. $PhysicalNames maps physical group ids to names. Version 4.1 restructures nodes and elements into per-entity blocks with a leading count line, which is why the two revisions must be parsed differently.

Specification at a glance#

sections	$MeshFormat, $PhysicalNames, $Nodes, $Elements, $Periodic
v2 nodes	tag x y z, one per line
v4.1 nodes	entity blocks: numBlocks, then per-block tags and coordinates
element codes	integer type ids (1 = 2-node line, 2 = triangle, 4 = tetrahedron, …)
groups	$PhysicalNames: dimension, tag, “name”

Read the full Gmsh specification ↗

Reading#

import polyxios as px

mesh = px.read("model.msh")
mesh.vertices          # (n, 3)
mesh.element_types     # one type code per element

Writing#

px.write(mesh, "out.msh")
px.write(mesh, "out.msh", float_fmt=".8g")

Format-specific options#

Option	Default	Effect
`float_fmt`	`".17g"`	On write: the format specifier for coordinates and data values. The default writes every float64 so it reads back unchanged.

Quirks worth knowing#

ASCII v2 and v4.1 are both read; writing emits v2.2, the revision every consumer still understands. A binary file, or any other version, is refused with CodecError.
Each element’s physical tag is read into element_attrs["phys_tag"], and the names $PhysicalNames gives them become element tags, so named boundaries and volumes survive the read. The writer takes phys_tag back when it holds one value per element, and derives the physical tags from element_tags otherwise.
Node tags need not be contiguous; they are remapped to dense indices.
$NodeData and $ElementData sections become vertex_attrs and element_attrs, at any component count the file declares, and are written back. A field is scattered by the tag each row names, so one covering part of the mesh lands where it belongs and the rest stays NaN. A row naming an entity the mesh does not hold - an element of a type this codec skipped, say - costs that row rather than the whole field.
A data section declares 1, 3 or 9 components and Gmsh refuses any other count, so a field of another width is padded out to the next of the three with zero columns and the padding is reported. The reader stays lenient and takes whatever width a file declares.
Gmsh numbers the mid-edge and face nodes of the higher-order elements by its own edge and face tables, which are not VTK’s; they are permuted on the way in and back on the way out. Gmsh’s 18-node prism becomes a biquadratic_quadratic_wedge.
Gmsh type 14, the 14-node pyramid, has no VTK equivalent and is skipped with a warning rather than reshaped into something else.
.msh is ANSYS Fluent’s extension as well, and the two are told apart by content: a file opening with $MeshFormat reads here, one opening with a parenthesised section reads as ANSYS Fluent. A bare write to .msh is Gmsh; pass fmt="fluent" for the other.