Philipp Rehner

With the lto profile being merged the performance gains change somewhat: model|components/segments|main (release)|main (lto)|new (release)|new (lto) -|-|-|-|-|- FMT|1|50.6 ms|28.2 ms|34.3 ms|**27.9 ms** PC-SAFT|1|68.3 ms|**27.6 ms**|53.1 ms| 30.1 ms PC-SAFT|2|205 ms|89.9...

Yes, but currently FeOs only determines the closest periodic image for every solid atom rather than every periodic image within the cutoff. When this is fixed, the check becomes obsolete.

Sounds good, I started the writing the guide using your suggested toc structure in #60

Very nice overview! Two quick comments without having gone through all the code: - If having a mixture of `SINumber` and `f64` is an issue, we can circumvent that by...

I added DFT benchmarks in #103. Comparing different solvers in a benchmark is not useful imo, because it gives a very limited insight into the capabilities of the solver. Do...

With the number of example notebooks that are already there, we should find a structure that makes the maintenance as simple as possible while making it easy to find the...

With the more lax trait bounds, the implementation can also be used to calculate FFTs of dual numbers (or [generalizations thereof](https://docs.rs/num-dual/latest/num_dual/)) that are used in forward automatic differentiation.

That would require redesigning around requiring the `Copy` trait bound that is implied by `Float`. Not impossible, but necessary? I would like to hear the maintainer's view on the issue...

Hi Iuri, fair point. The comment in the documentation only refers to pure components. I'm traveling right now without a laptop, so I didn't completely look through this, but if...

Due to the changes in https://github.com/PyO3/rust-numpy/pull/442, this crate does not build python packages for 32-bit Windows anymore.