Version History

Release series 0.2 - PyTorch integration

Fixed GPU memory leaks when using PyTorch back-end, thus enabling the solution of larger problems.
Improved documentation and automated deployment.

Implemented the ScatteringMatrix class, which allows calculation of scattering, reflection,and transmission matrices for any material for which the MacroMax solver can calculate the electro-magnetic field.
Added TensorFlow back-end, allowing access to, for instance, Google Colab’s Tensor Processing Units.

Streamlined iteration: 3x speed improvement of benchmark with PyFFTW on Intel Core i7-6700.
Reorganized parallel_ops backend.
Expanded unit testing.

Input arguments for isotropic materials or scalar calculations do not require singleton dimensions on the left anymore.
macromax.solve(...) and macromax.Solution(...) now take the optional input argument refractive_index as an alternative to the permittivity and permeability.
The macromax.bound module provides the Bound class and subclasses to more conveniently specify arbitrary absorbing or periodic boundaries. Custom implementations can be specified as a subclass of Bound.
Convenience class macromax.Grid provides an easy method to construct uniformly plaid sample grids and their Fourier-space counterparts.

The solve function and Solution constructor now take dtype as an argument. By setting dtype=np.complex64 instead of the default dtype=np.complex128, all calculations will be done in single precision. This only requires half the memory, and typically also reduces the calculation time by about 30%.

The current density can now be specified directly for solve(...) instead of the source_density.
Reorganized the file structure.
Removed the k0^2 = (2pi/wavelength)^2 factor again after the calculations. This normalization factor was introduced to avoid underflow in the iteration. Now it is removed as intended, after the iteration, or when an intermediate solution is queried.