·
12 commits
to develop
since this release
New features
- Pre- and postprocessing for elemental mixtures added
- MALA can now natively handle multielement systems (both with bispectrum and ACE descriptors, see below)
- This feature is still tested in production, so now extensive documentation exists at this point
- Splitting along the energy axis now supported
- This massively reduces the data cost for systems including core electrons
- ACE descriptors implemented, require specific LAMMPS version to function
- More details will be described in upcoming paper
- Currently, now GPU calculation available
- Descriptors can now be calculated on-the-fly
- This is especially useful in the GPU-accelerated case, since the GPU calculation of the descriptors is typically very fast and can be done directly before the training
- Through this functionality, descriptor related hyperparameters can be fixed more easily
Changes to API/user experience
- DataScaler are now saved in .json format
- MALA tested up to Python 3.13.3
Fixes
- Fixed bug that broke DDP performance
- Several smaller internal API fixes
- Improved CI