|
| 1 | +# Changelog |
| 2 | + |
| 3 | +All notable changes to ShiftML are documented in this file. |
| 4 | + |
| 5 | +## [0.2.0] - 2026-08-24 |
| 6 | + |
| 7 | +### Added |
| 8 | + |
| 9 | +- **New model `ShiftML4`.** Trained on PBE0 molecular-corrected GIPAW-PBE data |
| 10 | + (1.2 million chemical shieldings from 12600 organic crystals), covering the |
| 11 | + same 12 elements as `ShiftML3`: H, C, N, O, S, F, P, Cl, Na, Ca, Mg and K. |
| 12 | + Against a hold-out set it reaches isotropic shielding RMSEs of 0.40 ppm for |
| 13 | + $^{1}\text{H}$ and 2.22 ppm for $^{13}\text{C}$, compared to 0.42 ppm and |
| 14 | + 2.24 ppm for `ShiftML3` on the same structures computed at GIPAW-PBE level. |
| 15 | + Select it with `ShiftML("ShiftML4")`. |
| 16 | +- `shiftml.utils.loading.load_model`, which rebuilds a model's metatomic |
| 17 | + wrapper with the installed metatomic. This makes the published model files |
| 18 | + independent of the `metatomic-torch` release that exported them, so a new |
| 19 | + metatomic no longer requires re-uploading models or releasing ShiftML. |
| 20 | + |
| 21 | +### Changed |
| 22 | + |
| 23 | +- **Dependencies were narrowed and modernised.** `MetatomicCalculator` moved |
| 24 | + out of `metatomic-torch` into its own `metatomic-ase` package, and ShiftML |
| 25 | + follows it. `metatensor-learn` and `metatensor-operations` are no longer |
| 26 | + required at run time: they are only needed to *train* models, and everything |
| 27 | + the published models need from them is already compiled into the TorchScript |
| 28 | + archives. Dropping them removes the main source of version conflicts with |
| 29 | + other metatensor-ecosystem packages such as `metatrain` and `uPET`. |
| 30 | + |
| 31 | + | package | 0.1.1 | 0.2.0 | |
| 32 | + | --- | --- | --- | |
| 33 | + | `python` | `>=3.9` | `>=3.10` | |
| 34 | + | `metatensor-torch` | `>=0.7.6,<0.9` | `>=0.10,<0.11` | |
| 35 | + | `metatomic-torch` | `>=0.1.2,<0.2` | `>=0.1.17,<0.2` | |
| 36 | + | `metatomic-ase` | — | `>=0.1.1,<0.2` | |
| 37 | + | `metatensor-learn` | `>=0.3.2,<0.4` | removed | |
| 38 | + | `metatensor-operations` | `>=0.3.3,<0.4` | removed | |
| 39 | + |
| 40 | + |
| 41 | + |
| 42 | +### Migration notes |
| 43 | + |
| 44 | +- Requires Python 3.10 or newer. |
| 45 | +- If you pin `metatensor-torch` or `metatomic-torch`, update those pins; the |
| 46 | + new ranges are not compatible with the 0.1.1 ones. |
| 47 | + |
| 48 | +## [0.1.1] - 2026-02-17 |
| 49 | + |
| 50 | +- Retry Zenodo downloads that fail with HTTP 423. |
| 51 | +- Documentation and README fixes. |
| 52 | + |
| 53 | +## [0.1.0] - 2025-06-15 |
| 54 | + |
| 55 | +- First public release. |
| 56 | + |
| 57 | +[0.2.0]: https://github.com/lab-cosmo/ShiftML/releases/tag/v0.2.0 |
| 58 | +[0.1.1]: https://github.com/lab-cosmo/ShiftML/releases/tag/v0.1.1 |
| 59 | +[0.1.0]: https://github.com/lab-cosmo/ShiftML/releases/tag/v0.1.0 |
0 commit comments