v0.2.0

Added

• Added AIMNet2TorchSim, an optional TorchSim ModelInterface wrapper for static evaluation, optimization, molecular dynamics, and autobatched workloads via the Python 3.12+ torchsim extra (torch-sim-atomistic>=0.6,<0.7).
• Added TorchSim external/API documentation and runnable examples/ts_opt.py and examples/ts_opt_pbc.py scripts.
• Added dedicated CI coverage for the Sella optional extra.
• Added TorchSim CI coverage on Python 3.13.

Changed

• Split Sella tests out of the ASE-only CI lane.
• Updated the CodeQL security workflow to github/codeql-action v4.
• Clarified README installation guidance now that AIMNet core dependencies already include the GPU-accelerated nvalchemiops package.

Fixed

• Made ASE and PySisyphus calculator modules importable in docs builds even when optional dependencies are not installed.
• Marked the local Hugging Face metadata propagation test with the hf marker so the HF CI lane runs it.
• Clarified that aimnet2-rxn supports only net-neutral systems.
• Fixed the reaction-path Hessian example to avoid compile_model=True, which is incompatible with Hessian requests.
• Corrected PySisyphus unit conversion documentation from Bohr to Angstrom input conversion.
• Repaired malformed Markdown fences in the batch-processing tutorial.

Documentation

• Expanded API docs coverage for DataGroup, config helpers, AEV modules, and long-range modules.
• Added a public import inventory to the API overview.
• Added aimnet2-rxn to the README Hugging Face repo list, docs index, and pre-trained model changelog inventory.
• Aligned CUDA wheel examples on the CUDA 12.6 PyTorch index.
• Renamed the molecular dynamics NPT section to match ASE NPT rather than Berendsen.