A CAD kernel that returns exact gradients without differentiating its own code

Exact Gradients Through a BRep Kernel

A CAD kernel that returns exact gradients without differentiating its own code

Gradient-based design needs derivatives of mass, inertia, or simulation results with respect to parameters like hole and fillet radii. In a BRep kernel these parameters pass through Booleans, fillets, and tessellation, which make discrete choices. The authors freeze the tessellation's topology and differentiate node positions instead: surface nodes via dual-number evaluation of analytic surfaces, vertices and trim nodes via implicit differentiation, with tangency rows for fillet blends. The result is exact for the frozen mesh and extends to second derivatives, a transposed pass, and rigid-body adjoints with contact. Validation against finite differences found two errors, now fixed.

The seam differentiates the mesh the kernel produces at the current parameters, with segment counts and triangulation held fixed.
  1. ecto

    Hey HN,

    I submitted this thread a few days ago and woke up to it on the front page. If anyone would like to discuss the ideas, feel free to email me. I won't be here to argue any longer.

    My hope is that before you write your next comment, you consider the person on the other side of the screen.

    I'm doing my best. Maybe it isn't up to your standards, but it's what I have. Criticizing my use of AI is pissing in the wind. I'm on my last leg financially, have a family and little free time, and just want to make the world better with the few good ideas I have.

    I hope you have a nice day.

  2. JoeDohn

    My first thought when I read (skimmed mostly) the paper is PGO (Profile guide optimization), except instead of optimizing machine code based on runtime profiling, this would optimize geometry based on physical or engineering objectives ?

  3. YuechenLi

    So, the Claude English to human English translation:

    Ordinarily, if you want to adjust the parameters of a CAD part, say, a bracket, to meet a constraint like a particular mass or center of mass, the approach is to change every parameter a little, rebuild the geometry to see what physical properties changes, and repeat until you get the result you wanted.

    This is of course a very expensive operation because you have to iterate through every parameter on a part every single time for any optimization pass.

    So VCAD's approach is to essentially record how each of the mesh verticies connect to each other after the BRep tesselates, then calculate what part of the mesh are actually affected when a parameter change, and then use that to guide their optimizer to get the physical result they wanted, so they only need to change the relevant parameters they found that affects properties instead of doing the computational equivalent of banging your head against the keyboard and hoping something works.

More from this day

2026-11-16