DFTK dev 2025 Orsay

[niklasschmitz]

Let's use this issue to collect ideas for the upcoming internal dev meeting in Orsay, 03.09. - 05.09. . The main focus is on pushing the state of DFTK-core code on data structures, performance, interface redesign, testing, and some autodiff.

We are still crowd-sourcing ideas and will later vote for some goals. Here's a few points collected so far:

• PlaneWaveBasis performance
  • more efficient update constructors (changing the model, or changing the XC functional, without recomputing grids and symmetries)
  • cheaply increasing the Ecut
• General SCF infrastructure
  • Making scfres a struct
  • Common interface for ground state algorithms (SCF, direct minimization, ...)
  • Ways to run base SCFs only once and cache result.
  • Warm-starting scf with a previous scfres (possibly from a different geometry or model) -> try to re-use 1) solutions & tolerances, 2) data structures?
  • Higher-level heuristics OUTSIDE of SCF call (Autotry mixings etc)
• ForwardDiff
  • Making AD stresses work with general libxc functionals beyond DftFunctionals.jl (type promotion issues to be solved)
  • Fix symmetry detection and AD interplay.
• Testing infrastructure
  • Test refactoring: Run only the essential tests in Windows and Mac
  • Split full test suite into multiple sub-chunks (which can run in parallel)
  • Test randomness initialised by one global seed
• Go through DFTK issue list
• Handling of antiferromagnetic Systems (symmetry wrong ? SCF does not find AFM solution)
• (...)

Open to suggestions!

[antoine-levitt]

Nice features that could probably be done (at least the bulk of it) in three days: hybrids, ultrasofts, DFT+U

[mfherbst]

Thanks @niklasschmitz for getting this started.

I think we should try to put some emphasis on reducing some of the debt we acquired over the years (e.g. tests, SCF datastructure). Would be cool to make a concerted effort to reduce this a bit.

[mfherbst]

Some additional points from me:

• Ultrasoft pseudopotentials #1125, Ultrasoft Pseudopotentials #1092 : Impacts on SCF and datastructures, how to deal with impact on assumptions about FFT grid, look forward towards PAW
• Memory versus compute for pseudopotential projectors / non-local pseudopotentials

[antoine-levitt]

Also a fun thing to try there : sic an AI (Claude code seems to be popular) on the issue list and see what comes out of it.

[Technici4n]

Another idea: Make DFTK computations reproducible (i.e. use fixed random seeds). See also #1110

[Technici4n]

I suppose we can close this issue now. It was nice seeing you all!