The work presents a way to compute design gradients through vcad, an open-source boundary-representation kernel, without differentiating its Boolean and fillet algorithms. It freezes the tessellation at the current design, then differentiates mesh-node positions: analytic surface evaluation handles interior nodes, while implicit differentiation of surface equations handles vertices and trim nodes. Fillet sensitivities are obtained by adding tangency constraints, avoiding finite differences even when one radius moves many blend surfaces. The resulting gradients support mass properties, directional second derivatives, a transposed pass that computes sensitivities for many parameters efficiently, and a rigid-body trajectory adjoint with penalty contact. If a parameter change alters topology, the method reports an error rather than returning a derivative across the change.
Validation compares results with finite differences on the frozen mesh and, where topology permits, independently rebuilt models; the checks exposed and helped fix two errors on drilled, filleted parts. The derivative is exact for the frozen mesh, not for the continuous shape: polygonization makes a fillet-radius derivative 6.7% too large at 16 segments per circle, falling to 0.11% at 128. Plans are local and must be recaptured at each optimization step. Adaptive tessellation can change node counts as parameters vary, introducing small jumps and noise in line searches; genuine topology changes also require shrinking or rejecting a step. The implementation runs in double precision on CPU, and its contact and trajectory model is limited to frictionless ground-plane contact and a restricted set of joints and objectives.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.