Automatic differentiation for density functional theory in Julia.
This project is part of Google Summer of Code 2021. https://summerofcode.withgoogle.com/projects/#6407471502983168
Student: Niklas Schmitz
Mentors: Michael Herbst, Antoine Levitt, Dhairya Gandhi, Keno Fischer
Main Project Goal: Hellmann-Feynman stresses via ForwardDiff
JuliaMolSim/DFTK.jl#443
Main PR implementing the needed infrastructure for ForwardDiff compatibility:
(merged) Hellmann-Feynman stresses via ForwardDiff and custom rules #476
JuliaMolSim/DFTK.jl#476
Follow-up PRs:
(merged) Extend ForwardDiff fallback for SVector norm to handle multiple partials #488
JuliaMolSim/DFTK.jl#488
(merged) Improve _apply_plan type stability #494
JuliaMolSim/DFTK.jl#494
How to use: To the end user, the new feature of calculating stresses of a converged solution can now be accessed by calling stresses = compute_stresses(scfres). For a complete example, see the testcases in https://github.com/JuliaMolSim/DFTK.jl/blob/master/test/stresses.jl.
Other stretch goals:
(draft) ForwardDiff example of implicit differentiation beyond the Hellman-Feynman theorem, to calculate a dipole moment.
JuliaMolSim/DFTK.jl#520
(draft) Hellmann-Feynman derivatives using ChainRules and Zygote #519
JuliaMolSim/DFTK.jl#519
(merged) Add nondiff rules for one ones zero zeros #465
JuliaDiff/ChainRules.jl#465
(merged) (Fix #446) Widen _mulsubtrans!! type signature #447
JuliaDiff/ChainRules.jl#447
-
NLsolve general implicit differentiation rrule https://gist.github.com/niklasschmitz/b00223b9e9ba2a37ed09539a264bf423
-
linsolve implicit differentiation rrule https://github.com/niklasschmitz/DifferentiableDFTK/blob/main/sandbox/1D/1D.jl#L128-L140
-
Application example on a small differential equation (linsolve inside nlsolve) https://github.com/niklasschmitz/DifferentiableDFTK/blob/main/sandbox/1D/1D.jl
NLSolve.jl: Anderson instability example #273 JuliaNLSolvers/NLsolve.jl#273
ForwardDiff
- AbstractFFT rules (based on mcabbotts draft JuliaDiff/ForwardDiff.jl#495, fixed some bugs & perf penalties; should ideally be upstreamed)
- norm of SVec at zero: custom rule to comply with chainrules convention (and consistency with norm(Vector(x))) otherwise very pleasant overall and easy to debug