How it works
The ONNX graph is the denoiser; the DDIM loop, classifier-free guidance, and the evaluator all run in JavaScript.
Volume-fraction control is calibrated and reliable in roughly 0.37–0.57. Stiffness conditioning separates low vs high by about 2×.
Solidify removes tiny floating specks and fills lone pinholes after sampling, so cells read as solid shapes rather than scattered noise; ranking then favours connected, few-piece designs. Every number shown is re-measured on the cleaned cell.
Optimize stiffness (SIMP) refines each sample with a few iterations of inverse-homogenization topology optimization — density-filtered for a minimum length scale. This is the "generate-then-optimize" idea: diffusion proposes, topology optimization polishes into a clean, connected, mechanically optimal shape (Sigmund 1994; Andreassen et al. 2011). In standard mode it maximizes bulk modulus; in auxetic mode it minimizes C1122 − β(C1111+C2222) with enforced symmetry (Sigmund 1995) and is kept only if it actually lowers the measured Poisson's ratio — so it can deepen auxeticity but never harm it (a short polish can't create re-entrant geometry from scratch).
Physics guidance folds the same objective into denoising instead of after it (TopoDiff-style; Mazé & Ahmed 2023): over the last few DDIM steps the clean-cell estimate is nudged along the stiffness (or, in auxetic mode, the negative-Poisson) gradient — one FE solve per cell per step, volume-preserved. Unlike the post-hoc polish it keeps the batch diverse, but it is slower.
Auxetic targets (ν < 0): routed to a second DDPM — a specialist trained on 2,000 FEM-labeled rotating-squares cells, conditioned on [vf, ν]. It samples a double batch and ranks by the measured ν (the mixed model cannot switch geometry family from a conditioning channel — four retraining attempts confirmed this). In auxetic mode vf control is approximate. Every number shown is homogenized from the generated cell, never assumed.
Press Generate.