Yingbo Ma

This issue seems to be resolved on the latest MTK.

It doesn't need to be reorthorgonalized. We are using the modified Gram-Schmidt process. Also, if you want to add classical GS with DGKS reorthogonalization criterion [1]. There is one in...

It doesn't need to be reorthorgonalized. Modified Gram-Schmidt is stable.

Currently, `NonlinearSolve` doesn't work with Zygote due to `build_solution`. ```julia julia> using ModelingToolkit, Zygote julia> Zygote.gradient((u, p)->ModelingToolkit.StructuralTransformations.numerical_nlsolve((u,p)->hypot(u, p)-cos(u), u, p), 0.1, 0.2) ERROR: MethodError: no method matching ndims(::Tuple{Float64}) Closest candidates...

How is `DiffEqSensitivity` related? This is nonlinear system.

I derived frule and rrule here: https://gist.github.com/YingboMa/4e4496f828c6a3179004f6d0ca224d2a Someone just need to write a performant implementation of it.

For performance reasons, the adjoint for `numerical_nlsolve` should just be a dozen lines of code that's non-allocating.

Yeah, exactly.

Is this possible with SciMLBase?

Should we have SciMLBasicSolver.jl then?