================================================================================ TRELLIS-generated Jet-Engine Bracket โ€” Surface FEA Dataset (DeepJEB++) ================================================================================ Generated: 2026-06-07 Source compute: FEA host .85 (143.248.221.85) Cases: 15,360 deployable TRELLIS-generated brackets (deployable_final.txt) -------------------------------------------------------------------------------- FOLDER LAYOUT -------------------------------------------------------------------------------- trellis_fea_dataset/ 1_surface_meshes/ 15,360 input surface meshes (.obj), native ~50k verts (source: output_v2/_tri) 2_boundary_conditions/ 15,360 per-case BC (.npz): clamped bolt + loaded lug nodes 3_fea_fields/ 15,360 per-case surface FEA fields (.npz) deepjebpp_labels.csv per-case SCALAR labels incl. mass (15,360 rows) metadata.json machine-readable dataset spec (material/loads/units/schema) README.txt this file trellis_fea_percase.csv legacy labels WITHOUT mass (superseded by deepjebpp_labels.csv) File naming: .obj / .npz, e.g. 012-015-diag_xz_mm_IS02 (mesh, BC npz, field npz, and label CSV row share the same id.) -------------------------------------------------------------------------------- FEM SETUP (linear elasticity, static) -------------------------------------------------------------------------------- Material: E = 113800 MPa, nu = 0.342 (isotropic) Units: length mm, force N, moment N*mm, stress MPa, displacement mm Mesh: surface decimated to FEM_TARGET_VERTS = 25000, cleaned (pymeshfix), volume tetrahedralized with tetgen switches "pq1.2/10Y" BC: 4 bolt holes clamped (Dirichlet), load applied on the lug (bolt/lug auto-detected + registered per case) 4 LOAD CASES (name, type, direction, magnitude): ver force ( 0.000, 0.0, 1.000) 35,600 N hor force (-1.000, 0.0, 0.000) 37,800 N dia force (-0.669, 0.0, 0.743) 42,300 N tor moment ( 0.000, 1.0, 0.000) 565,000 N*mm -------------------------------------------------------------------------------- FIELD NPZ CONTENTS (numpy .npz, np.load) -------------------------------------------------------------------------------- Geometry (shared by all 4 loads): surface_points (N,3) float32 surface vertex coords (mm) surface_faces (M,3) int32 triangle indices R float detected lug bore radius (mm) cov_min, dXZ float lug-detection diagnostics Per load L in {ver, hor, dia, tor}: L_U (N,3) float32 surface displacement vector (mm) [for deformed viz] L_vm (N,) float32 surface von Mises stress (MPa) L_maxu float max displacement magnitude over the volume (mm) L_p95vm float 95th-percentile von Mises over volume nodes (MPa) -------------------------------------------------------------------------------- BOUNDARY CONDITION NPZ CONTENTS (2_boundary_conditions/.npz) -------------------------------------------------------------------------------- *** ALL indices are into the npz surface_points array of 3_fea_fields/.npz *** *** (the 25k FEM surface mesh, ~24.9k verts) โ€” NOT the native 50k meshes/.obj. *** bolt_idx (B,) int32 clamped (Dirichlet) surface nodes = the 4 bolt holes bolt_holes (B,) int8 hole label 0..3 for each bolt_idx node (spatial k=4; a few cases may show 3 labels if two holes merge) lug_idx (L,) int32 loaded (Neumann) surface nodes on the lug/clevis bore (same load nodes shared by all 4 load cases) n_surface int32 number of surface_points (sanity vs fields npz) R float32 lug bore radius (mm) jaccard_bolt float32 QC: agreement of re-detected vs U==0 bolt set (~1.0) frame str "surface_points (mm, 25k FEM mesh); same index as 3_fea_fields/.npz" Provenance: - bolt_idx recovered EXACTLY from the solved fields: nodes with U==0 in ALL 4 load cases (the Dirichlet-clamped set the solver actually fixed). - lug_idx re-detected with the deterministic ยง3.4.1 bolt/lug pipeline (trellis_local.prep) and mapped to surface_points via nearest-neighbour -- the same mapping fem_solver_nf uses; verified against the U==0 bolt set (mean Jaccard 0.9997 over all 15,360). Usage example: bc = np.load("2_boundary_conditions/.npz") fp = np.load("3_fea_fields/.npz")["surface_points"] clamped_xyz = fp[bc["bolt_idx"]] # mm coords of clamped bolt nodes loaded_xyz = fp[bc["lug_idx"]] # mm coords of loaded lug nodes -------------------------------------------------------------------------------- SCALAR LABEL COLUMNS (deepjebpp_labels.csv) -------------------------------------------------------------------------------- case, mass_g, vol_mm3, cov_min, dXZ, R, ver_maxu, ver_p95vm, ver_maxvm, hor_maxu, hor_p95vm, hor_maxvm, dia_maxu, dia_p95vm, dia_maxvm, tor_maxu, tor_p95vm, tor_maxvm - mass_g = mass (Ti-6Al-4V, rho=4.43e-3 g/mm^3) x enclosed volume of the registered native ~50k mesh (full-resolution, pymeshfix-watertight). - vol_mm3 = that enclosed volume (mm^3). - The 3_fea_fields npz store maxu/p95vm; the CSV additionally has maxvm = max von Mises. (trellis_fea_percase.csv is the same table WITHOUT mass_g/vol_mm3 โ€” superseded.) -------------------------------------------------------------------------------- NOTES -------------------------------------------------------------------------------- * All 15,360 fields are at the SAME 25k mesh density (consistent labels). * 6 geometrically complex cases failed the standard 25k decimation (self- intersecting facets -> tetgen reject). They were re-solved with a robust decimation (pymeshfix-clean the original BEFORE decimating; same 25k target and same tetgen switches). Their scalars match the CSV within ~0.1-18% (vs +30-170% if solved at 200k), so they remain on the 25k scale: 187-384-diag_xz_mm_IS06, 217-136-diag_xz_mm_IS09, 384-123-diag_xz_mm_IS00, 384-136-diag_xz_mm_IS18, 384-279-diag_xz_mm_IS13, 384-548-diag_xz_mm_IS08 * Field npz store surface (boundary) values; max/p95 scalars are over the full volume mesh. ================================================================================ -------------------------------------------------------------------------------- LICENSE -------------------------------------------------------------------------------- This dataset is distributed under the Open Data Commons Attribution License (ODC-By) v1.0. See the LICENSE file. Free to use, modify, and share -- including commercially -- with attribution (please cite the DeepJEB++ paper and the upstream SimJEB and DeepJEB datasets). https://opendatacommons.org/licenses/by/1-0/