Auto3D 3.5.0

Breaking Changes

• Python 3.11+ and PyTorch 2.8+ required - Dropped support for Python 3.10
  and PyTorch < 2.8. torchani now requires >= 2.8 for the ANI2x/ANI2xt engines.

• Bundled AIMNet2 .jpt models removed - AIMNet2 is now provided by the
  aimnet package (a core dependency) instead of files shipped inside Auto3D.
  Models are auto-downloaded and sha256-validated into ~/.cache/aimnet on
  first use; set AIMNET_CACHE_DIR to override the cache location. Network
  access is required once per model.

• Default AIMNet energies and forces differ from 3.x - The aimnet
  registry .pt externalizes D3 dispersion (vs the embedded-D3 .jpt used in
  3.x). Absolute E_tot values shift and conformer rankings may differ
  slightly as a result.

• Thermo entropy property renamed - The thermochemistry output SDF property
  S_hartree is now S_hartree_per_K, correctly reflecting its units
  (Hartree/Kelvin). Update any code that reads the old property name.

Added

• Registry model selection - optimizing_engine now accepts any aimnet
  registry name (aimnet2, aimnet2-2025, aimnet2-nse, aimnet2-pd, ...) and
  custom model file paths, in addition to AIMNET, ANI2x, and ANI2xt.
  AIMNET remains an alias for the registry default aimnet2.

• CLI surfaces the registry - auto3d models list now shows the AIMNet2
  registry families, and auto3d models info reports the correct 14-element
  AIMNet2 set (H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I).

• First-class property subcommands - auto3d energy, auto3d optimize,
  auto3d thermo, and auto3d tautomers expose single-point energy, geometry
  optimization, thermochemistry, and tautomer ranking from the CLI (previously
  Python-only). Each supports --engine, --gpu/--no-gpu, --gpu-idx,
  -o/--output, and --json.

• auto3d models test <engine> - loads an engine and runs a tiny forward
  pass as a health check (catches a missing torchani, a failed aimnet registry
  download, or a broken custom model file before a full run).

• Live optimization progress - interactive auto3d run shows a live panel
  (converged / active / dropped / step) during geometry optimization instead of
  a static spinner.

• CLI ergonomics - auto3d run gains --job-name and --save-intermediate;
  config init gains --force; choice flags use enums with shell completion;
  input paths are validated up front; and commands return differentiated exit
  codes (2 config/input, 3 dependency, 4 GPU, 5 model).

• API additions (backwards-compatible) - calc_spe, opt_geometry, and
  calc_thermo accept out_path, use_gpu, and allow_tf32. A canonical
  generate_conformers alias for main() is exported (main still works), and
  get_stable_tautomers / select_tautomers are now part of the public API
  (Auto3D.__all__). main() returns a WorkflowResult -- a str subclass
  holding the output SDF path (drop-in for the previous return) plus
  n_molecules / n_conformers counts.

Changed

• use_ensemble no longer loads a bundled 8-model ensemble file. A single
  registry member is used; passing use_ensemble=True now emits a warning.

• allow_tf32 now applies to the energy/optimize/thermo paths. These
  previously selected the device inline and ignored TF32; they now route through
  the shared device + torch configuration, so enabling TF32 affects them too
  (a small numerical change for anyone who had set it expecting it to apply).

• Thermochemistry reference temperature is now 298.15 K - The default
  temperature for thermodynamic property calculation changed from 298 K to the
  standard 298.15 K.

Fixed

• smiles2mols() no longer silently drops inputs that share an InChIKey -
  Distinct inputs that collapse to the same standard InChIKey (e.g. some
  tautomers, or the same molecule written two ways) are now disambiguated with a
  suffixed id and a log message instead of being dropped.
• Energy-guarded conformer deduplication - Conformers within the heavy-atom
  RMSD threshold are merged only when their energies also agree within
  DEFAULT_DUPLICATE_ENERGY_TOL, so genuine O-H/N-H rotamers are no longer
  collapsed.
• Thermochemistry robustness - Spin multiplicity is derived from the
  molecule's radical electrons (with a warning that NNP energies are
  closed-shell), and imaginary vibrational modes are ignored rather than
  failing the whole molecule.
• GPU index is validated up front - An out-of-range gpu_idx now raises a
  clear configuration error instead of crashing inside a worker; a CPU run with
  a list of GPU indices no longer spawns redundant contending workers.
• CLI reports a real failure count - auto3d run reports the number of
  input molecules that produced no conformer instead of always reporting zero.
• Deterministic file handling - combine_smi preserves input order, and
  .smi molecule indexing is gap-free when blank lines are present.
• Read the Docs build - the docs build now runs on Python 3.11 to match
  requires-python, fixing the failing hosted documentation build.

Removed

• Dead torch.jit.optimized_execution guard in the batch optimizer (a no-op for
  the eager-mode model wrapper).