Skip to content

CHANGELOG.md

Latest commit

 

History

History
148 lines (131 loc) · 7.68 KB

CHANGELOG.md

File metadata and controls

148 lines (131 loc) · 7.68 KB
Raw

Changelog

All notable changes to MLatom are summarized here. For the complete, detailed release notes see the official release notes.

Dates are given as DD.MM.YYYY. Versions are available on PyPI and GitHub.

[3.25.2] – 20.08.2026

  • AIQM3@DFT* can now be requested: MLatom did not recognize it as a method. By Pavlo O. Dral.
  • Fixed: reading Gaussian or ORCA output could fail with No module named 'rmsd' on a fresh install. rmsd is now installed together with MLatom. By Pavlo O. Dral.
  • Fixed: UV/Vis spectra lost the peaks they had computed when IPython was not installed. By Pavlo O. Dral.
  • Fixed: MLatomF could hang on long runs, and its error messages were not shown. By Pavlo O. Dral.
  • A method that needs a program you do not have now names that program and gives the command that installs it. By Pavlo O. Dral.
  • Molecular-orbital energies and occupations are available from the PySCF, xTB, Gaussian and ORCA interfaces. By Pavlo O. Dral.

[3.25.1] – 14.08.2026

  • Fixed: the ANI -D4 models still selected their dispersion parameters by functional name, so on dftd4 4.0.0 and newer they used a different fit — 3.9 kcal/mol on an ethanol total energy, silently. 3.25.0 fixed this for AIQM1, AIQM2 and OMNI-P1 but not for these. No method now depends on how dftd4 resolves a functional name. By Pavlo O. Dral.
  • Fixed: import mlatom required PyTorch to be installed, because aiqm1.py imported it at module level without using it. By Pavlo O. Dral.
  • AIQM3's D3(BJ) term is given explicitly too. Results are unchanged. By Pavlo O. Dral.
  • We still recommend dftd4 3.6.0: that is the version MLatom is tested against. The parameter pinning above makes the D4 energies identical on 4.x, but the test suite has not been run against 4.x, so Hessians, thermochemistry and the rest are unverified there.

[3.25.0] – 13.08.2026

  • Fine-tuning of the universal models on your own data. ANI, AIQM1, AIQM2, AIQM3, UAIQM, OMNI-P1 and OMNI-P2x are fine-tuned through the same train() call, and the result saves and loads like any other model. See the tutorial.
  • Dispersion is now declared once and handled on both sides: the term is subtracted from the reference labels before training and added back at prediction, so a fine-tuned model keeps the long-range behaviour of the model it started from.
  • One geometry can carry labels from several methods. A molecule records where a label came from (molecule.label_source), a property can be addressed by its source (mol.get_property('wb97x.energy_gradients')), and the database keeps a registry of the sources it holds (molecular_database.label_sources). Databases can be written to and read back from HDF5.
  • Molecules missing the property being learned are dropped before training rather than entering the loss as NaN, for the training and the validation set alike (molecular_database.without_missing_labels()).
  • Fixed: a weighted_sum model tree applied its weights to only one molecule of a database, so every multi-molecule weighted sum was in effect unweighted. This affected all DENS predictions.
  • Fixed: the D4(wB97X) term of AIQM1, AIQM2, OMNI-P1 and the ANI -D4 variants is now selected by its damping parameters rather than by the functional name. dftd4 4.0.0 renamed wb97x to wb97x-2008 and gave the name wb97x to a different functional, so on dftd4 >= 4.0.0 these models silently used another functional's damping: AIQM2 shifted by 3.5 kcal/mol on one geometry, and the methane-dimer well deepened from -0.65 to -1.16 kcal/mol. Results on dftd4 3.x are unchanged.
  • Fixed: re-downloading a model died on the files already present, leaving the model half-extracted.
  • Fixed: CISD and CCSD were concatenated in the PySCF interface's list of supported methods, which removed both from dispatch.

[3.24.0] – 03.08.2026

  • MLatom now prints the references to cite for AIQM3, UAIQM, OMNI-P2x, ANI-1ccx-gelu, DFT ensembles, and for IR spectra with AIQM models.
  • Fixed the state energies read from MNDO's fort.15 when the requested gradient list does not start at the ground state. This affected the energies reported along a molecular dynamics trajectory, not the trajectory itself.
  • Fixed the filtering of initial conditions, where the molecules kept in the filtered database were all the same object.
  • Downloading a model no longer hangs indefinitely on an unresponsive server, and a download that did not succeed is recognized as such: MLatom continues with the next link and, if none work, reports where to get the files by hand.
  • Fixed the type annotation of molecule.nstates, which used np.int, removed in NumPy 1.24.

[3.23.5] – 27.07.2026

  • Sampling now writes the training/subtraining/validation/test and cross-validation index files correctly to output paths that contain spaces (and portably across platforms), replacing shell cp/mv calls with shutil. Thanks to @rayair250-droid.

[3.23.4] – 15.07.2026

  • MLatom is now released under the Apache License 2.0.
  • Improved initial-conditions sampling: unified random and Maxwell-Boltzmann velocity generators, with correct angular-momentum removal and linearity-aware degrees of freedom for linear molecules and reproducible sampling.
  • Retuned and hardened the state-gap loss used in OMNI-P2x fine-tuning and multi-state active learning.
  • pip install mlatom now installs all required dependencies automatically — no manual dependency list needed.
  • Fixed a crash when a method was queried while PySCF was not installed.
  • The package can now be installed directly from a source clone (pyproject.toml).

3.23.3 – 27.06.2026

3.23.2 – 26.06.2026

3.23.1 – 25.06.2026

  • Bug fixes and performance improvements, including faster NAMD.

3.23.0 – 15.06.2026

  • AIQM3 is now available as a public add-on (pip install aitomic-addons).
  • Startup banner now reports the running version together with its commit and build date; MLatom checks once a day whether a newer release is available.
  • Direct Gaussian workflows: MLatom generates Gaussian input and runs Gaussian for Gaussian-internal methods combined with geometry optimizations, frequencies, IRC, and excited-state (TD) calculations.
  • Orientation fixes for normal modes and dipole moments.
  • Refined ORCA 6 interface (dispersion and ground-state energy parsing).
  • Bug fixes across MNDO gradients, FSSH non-adiabatic couplings, NAMD, dihedral-angle handling, and active learning.

3.22.0 – 09.03.2026

  • Released OMNI-P2x with tutorials.
  • Improved performance of NAMD simulations.

3.21.0 – 13.02.2026

  • Refactored ORCA interface supporting many more excited-state simulations (e.g., QD-NEVPT2).
  • Quality-of-life improvements for analyzing UV/vis absorption spectra.

3.20.0 – 26.12.2025

  • TDBA and other improvements in FSSH.

[3.19.0 – 3.19.1] – 23.10.2025 / 14.11.2025

  • See the full release notes.

[3.18.0]

  • FSSH; KRR in Julia; MDtrajNet-1 (universal model for directly predicting MD trajectories); ECTS (a diffusion model for generating transition states).

Earlier releases

For 3.0.0 through 3.17.x and detailed per-version notes, see the official release notes.