mcsnekkeh/scriptcad-carve

ScriptCAD Carve: 3D topology optimisation, with engineering outcomes

851 parts carved to their stiffest shape for their weight by ScriptCAD's Carve tool, each with its own dimensions, supports, load, material budget and smallest member size:

Family	Samples
random	211	a base block with up to three blocks and cylinders on it, on feet or a wall, one or two load pads, load at a random angle
flange	100	a plate with 2-6 bolt holes (held in the holes) and a loaded boss
lbracket	99	an L held at the back of its upright, loaded at the end of its arm
chair	148	the lounge-chair block: side profile, 3 or 4 feet, seat and back loaded, optionally rounded
beam	149	a bar on two or three feet, or held at one end, with a load pad on top
bracket	144	a block bolted to a wall by two or four pads, with a load pad on top or in front

Every design comes from a full optimisation in 3D, with every step solved by finite elements; nothing is predicted. Each part is then labelled the way an engineer would judge it:

• Would it hold? A real alloy and process, a load that gives the uncarved design space a realistic margin, and the carved part's stress, safety factor and deflection under it. 562 of 851 (66%) pass both a safety factor of 1.5 and a deflection of at most 1/500 of the span; the rest are labelled as failing, which is as useful to learn from.
• Can it be made? The number of pieces, the thinnest member, and for 3D printing the unsupported overhang for each of the six build directions.
• Is the result right? Each part is rebuilt from its own CAD document, and its carved design re-solved independently in float64.
• What is the recipe? The parametric ScriptCAD document, the same model as Python code (the kernel's own), and a plain-language spec: spec → CAD code → design space → carved design, in one row.

The carved designs are on median 2.45× (up to 10.05×) as stiff per gram as their design space. 49 drawn parts failed to set up (their load face was not found); they are kept, marked status: failed, with the reason, and have no design.

Open any part in ScriptCAD (save scriptcad_doc as part.scad.json, File → Open project) or in the online demo, change it, and carve it again.

What is in a row

The viewer shows a preview of each design: the view along its thinnest side, the other side view, and the top view (lighter is nearer; the design space is a faint shadow).

Column
status, failure	ok, or failed with the reason
family, params	the part family and its parameters (JSON)
spec	the requirement in plain language
material, process, E_MPa, poisson, density_g_cm3, yield_MPa	a real alloy and how it would be made
load_N, fixed_points_mm, load_points_mm	the total force (spread over the load faces) and a point on each support and load face
load_basis	where the load comes from: a seated person for chairs, otherwise scaled to the design space's margin
keep_target, keep_reached, keep_floor	the material share asked for, reached, and the share that supports and loads keep on their own
detail	smallest member: fine, medium, coarse (about 2.5, 4 and 6 voxels at Normal resolution)
mass_g_before, mass_g_after	mass of the design space and of the carved part
safety_factor_before, safety_factor_after	yield over member stress, uncarved and carved
member_stress_MPa_before, member_stress_MPa_after	99th percentile of von Mises away from the supports and load faces
corner_peak_MPa_after	the single highest stress away from them (at voxel-sharp corners; a real part has fillets there)
max_disp_mm_before, max_disp_mm_after, deflection_limit_mm	largest displacement, and span / 500
passes_strength, passes_deflection, passes	safety factor ≥ 1.5; deflection within the limit; both
stiffness_per_mass_gain	(compliance × mass) of the design space over that of the carved part
pieces, largest_piece_share	connected pieces of the design, and the largest one's share
thinnest_member_mm, member_mm_median	member thickness (5th percentile and median along the medial voxels)
print_direction, print_overhang_unsupported, overhang_by_direction	for 3D printing: the build direction with the least unsupported overhang (45° rule), that share, and all six (JSON)
reproducible, f64_rel_diff	rebuilt from its document to the same grid; float64 re-solve vs the optimiser's float32 compliance
compliance_before, compliance_after, carve_load_N, carve_E_MPa	the carve as it ran (its own load and stiffness)
voxel_mm, grid_xyz, origin_mm	the voxel grid
solves, seconds, stop, history	how the optimisation went (history: share, compliance and deflection per step, JSON)
scriptcad_doc, python	the part as a ScriptCAD document and as Python for its kernel
arrays	the path of the voxel arrays

arrays/<id>.npz holds the voxel arrays, indexed [z, y, x]:

• fill (uint8, 0–255): how full each voxel of the design space is (surface voxels are partly full)
• design (bool): the carved design
• held_nodes (bool, one larger in each direction): the nodes the solver holds. These are the supports, and also every node that touches no voxel of the design space. The supports are the held nodes that touch at least one voxel with fill > 0
• node (I, J, K) has index n = I + (nx + 1) * (J + (ny + 1) * K), so held_nodes[K, J, I]; degree of freedom 3 * n + axis, axis 0, 1, 2 = x, y, z
• force_dofs, force_values: the loaded degrees of freedom and their forces; they sum to carve_load_N

from huggingface_hub import hf_hub_download
from datasets import load_dataset
import numpy as np

ds = load_dataset("mcsnekkeh/scriptcad-carve", split="train").filter(lambda r: r["status"] == "ok")
row = ds[0]
print(row["spec"])
z = np.load(hf_hub_download("mcsnekkeh/scriptcad-carve", row["arrays"], repo_type="dataset"))
design, space = z["design"], z["fill"] > 0           # [z, y, x]
print(row["family"], design.shape, row["safety_factor_after"], row["passes"])

How it was made

bench/dataset/ in the ScriptCAD repository (generate.py, then label.py, then pack.py):

1. Draw the part. A family generator draws the dimensions, supports, load direction and carve settings from a seeded random generator (sample i uses seed 1000 + i).
2. Build and voxelise. The CAD kernel (NVIDIA's open-source SMLib, USD BRep) builds the part, and ScriptCAD voxelises it at Normal resolution (about 200,000 voxels), with partly full voxels at the surface.
3. Carve. The app's own optimiser (hard-kill BESO, smoothed over the smallest member size) removes the material that stores the least strain energy, 3% per step, until the share asked for is left and the stiffness has settled. Material at supports and loads always stays, and parts that lose their support are removed. Every step is a linear-elastic finite element solve (8-node hexahedra with incompatible modes) on the GPU (CUDA multigrid, float32).
4. Label.
   • The part is rebuilt from its document, and the design space and the carved design are solved in float64.
   • A material is drawn with the Poisson's ratio the part was carved with. In linear elasticity the optimal layout depends only on the load's direction and on Poisson's ratio, not on the material's stiffness or the load's size, so the design is the right one for it.
   • Chairs carry a seated person: 2.1 kN in total (about 1.6 kN on the seat and 0.5 kN on the back, as in furniture static strength tests). The safety factor follows from it, and shows how far a chair could be carved further.
   • Other parts have no known load, so theirs is scaled to give the design space a safety factor drawn from 3 to 15 (log-uniform): the designer gave the space margin, and Carve spends it.

Limits

• One load case per part; linear elasticity; strength judged by yield on member stress (no fatigue, buckling or contact).
• Designs are voxel grids. Supports and loads act on whole faces, held rigidly. The seat and back of the chairs, being loaded faces, stay as plates.
• Hard-kill BESO gives designs within 1–2% of the stiffness of a SIMP reference on the MBB beam when members span at least about 3 voxels (benchmark).
• Design spaces are generated, not taken from real products; half of the newer samples are random shapes to cover more ground. Real-part design spaces, keep-out and keep-solid regions, and several load cases per part are planned.

Licence and citation

CC BY 4.0. Please cite ScriptCAD: https://github.com/rubenhekkens/scriptcad.