Skip to content

MLatom 3.26.0

Choose a tag to compare

@dralgroup dralgroup released this 07 Nov 08:39
· 2 commits to main since this release

MLatom 3.26.0 — released 7 November 2026

  • New: the universal MACE models for materials, MACE-MP-0 and MACE-MPA-0, support periodic cells. molecule.pbc and molecule.cell now reach every MACE model, and a periodic molecule without a cell is refused instead of giving a meaningless energy. Their weights are not shipped: download them before the first use and place them in ~/.mlatom/models/mace_mp_model/ (or $MODELSPATH/mace_mp_model/), as for the MACE-OFF models. By Pavlo O. Dral.

  • Fixed: MACE-OFF calculations failed with a mixed-precision error. MLatom now holds the model's precision while it predicts, and loading a model again keeps the requested precision. By Pavlo O. Dral.

  • Fixed: analytic Hessians of KREG models trained on gradients omitted the gradient-trained contribution. These Hessians now use batched central differences of KREG's analytic gradients. By Pavlo O. Dral and Yifan Hou.

  • Changed: numerical gradients use central differences by default, with the old forward route still available. Numerical Hessians use central second differences of energies by default; central differences of model-returned gradients are available explicitly. Sparrow's numerical derivative settings can be configured per calculation. By Pavlo O. Dral and Yifan Hou.

  • Fixed: analytic Hessians of ANI-type models could be wrong when two atoms had nearly equal coordinates. Energies are unchanged; gradients of ANI-type models change at the 1e-9 Hartree/Å level, so geometry searches can take a different number of steps. By Pavlo O. Dral.

  • Fixed: built-in IRC with hess_est=False stopped with AttributeError instead of calculating an exact Hessian at every step as documented. By Pavlo O. Dral.

  • Fixed: ml_program='mlatom.jl' could fail on its first import when Python was statically linked to libpython. MLatom now retries with Julia's module cache disabled. By Pavlo O. Dral.

  • Fixed: ANI and MS-ANI training used ReduceLROnPlateau.is_better, which newer PyTorch versions made private. Both the old and new method names are now supported. By Pavlo O. Dral.

  • Fixed: gap_md uses the requested gap_threshold instead of always using 0.03. By Mikolaj Martyka.

  • Fixed: ANI-1xnr looks for its parameters in $MODELSPATH/ani_1xnr_model/ before falling back to the packaged defaults. By Pavlo O. Dral.

Known issues

  • The native MLatomF kernel solver can return incorrect results when Cholesky fails and its fallback is used; this was observed for periodic kernels. Repair is deferred.
  • MDtrajNet propagation using numerical velocities can produce different trajectories across platforms. A portable production repair is deferred.