mapengfei/dalton-lv-meshes

Dalton et al. (2023) Left-Ventricle Meshes — DOLFINx XDMF

Finite-element meshes of 2967 synthetic left-ventricle (LV) geometries, converted from the raw node data of

D. Dalton, H. Gao, D. Husmeier, "Emulation of cardiac mechanics using Graph Neural Networks", Computer Methods in Applied Mechanics and Engineering 401 (2023) 115645. Dataset: Zenodo 7075055 · Code: https://github.com/dodaltuin/passive-lv-gnn-emul

The upstream download contains only a node-adjacency graph plus per-geometry node coordinates/features — no element connectivity. This dataset adds the reconstructed elements:

• the shared topology is a 3-cell-thick hexahedral shell (6784 nodes, 4995 hexes, 4 through-thickness nodal layers of 1696 nodes);
• every hex is split into 12 tetrahedra (centroid fan with globally consistent face diagonals) → 59 940 conforming P1 tetrahedra, 11 779 nodes per mesh.

All 2967 meshes share the identical topology; only the nodal coordinates and the per-geometry fibre/sheet features differ.

Layout

.
├── geom_0000/   dalton_lv_mesh.xdmf  dalton_lv_mesh.h5  meta.json
├── geom_0001/   ...
│   ...
├── geom_2966/
├── manifest.csv            # geom, split, local_index, nodes, tets, volume, base_z, path
└── upload_to_hf.py

Global index → source split:

geom range	split	local indices
geom_0000 … geom_2249	train	0 … 2249
geom_2250 … geom_2399	validation	0 … 149
geom_2400 … geom_2966	test	0 … 566

Contents of each mesh (dalton_lv_mesh.xdmf + .h5)

item	description
topology	P1 tetrahedra, cell type tetrahedron, 4 nodes/cell
cell_tags	1 = MYOCARDIUM
facet_tags	5 = BASE, 6 = ENDO, 7 = EPI
fibers	per-cell transmural fibre direction f0 (Vector, Center="Cell")
sheets	per-cell sheetlet direction s0
normals	per-cell n0 = f0 × s0

meta.json records the global id, source split/local index, wall volume (cm³) and the base-plane coordinate. Units are cm.

Coordinates are rotated to a fixed-base convention x' = z − z_base, y' = y, z' = −x, so the flat base plane lies at x' = 0 (matching the authors' FDM/wave pipeline).

Fibres: f0 from the upstream node features (cols 5:8), sheet angle from cols 8:9, s0 = cos β (e_r × f0) + sin β e_r with e_r the through-thickness chain direction; per-cell values are the P1 centroid average followed by Gram-Schmidt.

Quick start (DOLFINx)

from mpi4py import MPI
from dolfinx.io import XDMFFile

with XDMFFile(MPI.COMM_WORLD, "geom_0000/dalton_lv_mesh.xdmf", "r") as xdmf:
    mesh = xdmf.read_mesh(name="mesh")
    mesh.topology.create_connectivity(mesh.topology.dim - 1, mesh.topology.dim)
    cell_tags  = xdmf.read_meshtags(mesh, name="cell_tags")
    facet_tags = xdmf.read_meshtags(mesh, name="facet_tags")

In ParaView, open the .xdmf (keep the .h5 in the same folder) and colour by facet_tags or fibers.

Validation

• every triangular face is shared by exactly 1 or 2 tetrahedra (conforming);
• all tetrahedron volumes are positive (no inverted cells);
• the facet tags sum to 7020 boundary triangles (ENDO 3330, EPI 3330, BASE 360);
• fibers/sheets/normals are orthonormal and right-handed (det = +1);
• DOLFINx re-assembles the same wall volume as meta.json (volume range 21.4 – 176.3 cm³, mean ≈ 82.5 cm³).

Citation

@article{dalton2023emulation,
  title   = {Emulation of cardiac mechanics using Graph Neural Networks},
  author  = {Dalton, David and Gao, Hao and Husmeier, Dirk},
  journal = {Computer Methods in Applied Mechanics and Engineering},
  volume  = {401},
  pages   = {115645},
  year    = {2023},
  doi     = {10.1016/j.cma.2022.115645}
}

The underlying simulation data are released under CC BY 4.0 by the original authors; please cite the paper and the Zenodo record. The reconstruction/conversion scripts are provided as-is.