loganbvh/superscreen: v0.9.0

SuperScreen v0.9.0 is a significant update indented to improve the CPU/memory efficiency and ergonomics of the package. The major upgrades are contained in the following Pull Requests: https://github.com/loganbvh/superscreen/pull/96, https://github.com/loganbvh/superscreen/pull/99.

1. Rather than creating a single mesh containing all films in a device, each film now gets its own mesh and finite element operators.

   • This approach avoids meshing large vacuum regions between films, which is costly.
   • This approach also means that films with transport terminals can exist in the same superscreen.Device as films without terminals. As a result, the superscreen.TransportDevice class is no longer needed and has been removed.
   • It is no longer necessary to explicitly define a bounding box around the film(s) being modeled. A suitable bounding box is automatically generated for each film within superscreen.Device.make_mesh().

2. Reduced memory footprint and improved performance using numba JIT compiled functions.

   • Several costly numerical operations have been migrated from numpy/scipy to custom numba just-in-time (JIT) compiled functions. In particular, scipy.spatial.distance.cdist(), which computes a matrix of pairwise distances between points in space, has been replaced with superscreen.distance.cdist(), written using numba. The numba implementation is automatically performed in parallel on multiple CPU cores and avoids the allocation of large intermediate arrays, which can cause significant memory usage in numpy/scipy.
   • For devices with multiple films, the inductive coupling between films is now calculated using the supercurrent density and a numba implementation of the Biot-Savart law, rather than the stream function and Ampere's law. The new approach is both more robust for closely-stacked films and avoids storage of large temporary arrays.
   • The default for superscreen.Device.solve_dtype has been changed from float64 to float32.

3. The linear system describing the magnetic response of each film is now LU factored only once per call to superscreen.solve().

   • This dramatically speeds up self-consistent simulations involving multiple films and makes the solver code more modular.
   • The portions of the model that are independent of the applied field can be pre-factorized using superscreen.factorize_model(), which returns a superscreen.FactorizedModel object that can be saved for future use. A superscreen.FactorizedModel instance can be passed directly to superscreen.solve().

4. As a result of the above optimizations, GPU support using jax and parallel processing with shared memory using ray no longer seem to add much value to the package, so they have been removed.

   • The gpu argument to superscreen.solve() has been removed, as has the (optional) dependency on jax
   • superscreen.solve_many() has been removed, as has the dependency on ray

5. All IO operations, including writing superscreen.Device and superscreen.Solution objects to disk, are now performed using the HDF5 file format via h5py.

   • All objects within superscreen that can be serialized to disk now have .to_hdf5() and .from_hdf5() methods.

6. SuperScreen has dropped support for Python 3.7, which will reach end-of-life in June 2023.

   • Added support for Python 3.11, which was being blocked by the dependency on ray.