Michael F. Herbst

> it might explain why the results are different ... and why our SCF converges faster. The collinear cases are usually much easier to converge > We do something similar,...

So it's now 16s (DFTK) versus 8s (QE) per SCF step, did I get that right?

> But indeed it's not far off. Also happy that we arrive to that conclusion. Anything else would have worried me quite a bit.

@antoine-levitt Didn't we try something like this in the LOBPCG at some point and discarded it, because it led to numerical instabilities, or do I misremember?

On that note I also thought about integrating with [TimerOutputs.jl](https://github.com/KristofferC/TimerOutputs.jl), which is kind of an optional light-weight profiling.

No, but I'll run it quickly.

So on my experiments (current master) I got: ``` DFTK: [150, 40, 40] ABINIT: 144 36 36 ```

current master PR not merged.

For reference on the point Antoine mentioned, you can reduce the cutoff for the points where XC quantities are computed using ```julia model_DFT(lattice, atoms, positions, Xc(functionals; potential_threshold=tol); kwargs...) ``` where...

Yes I have seen that, thanks. For sure the kpath and BZ plot might be very nice to expose. Perhaps also to integrate with you closer for the band structure...