If you’ve got some money,

https://www.comsol.com/blogs/6-examples-of-simulation-driven...

Those examples wouldn’t be differentiable in the sense of automatic differentiation. “Just” parametric CAD + FEM/BEM + optimizer.

Though differentiable FEM has kind of been done for a while under the name “adjoint method” for topology/shape optimization. It’s not super common outside of academia though.

Personally, I’m keeping an eye out for projects like https://github.com/deepmodeling/jax-fem to make some waves in that space.