v0.1.1

What's Changed

Breaking Changes

• Minimum PyTorch raised from 2.4 to 2.8
• Minimum nvalchemi-toolkit-ops raised from 0.2 to 0.3
• Creating new TorchScript modules via torch.jit.script() is no longer supported (legacy .jpt loading preserved)

Changed

• Modernized nvalchemiops imports to v0.3 API (nvalchemiops.torch.neighbors, nvalchemiops.torch.interactions.dispersion)
• Replaced deprecated torch.inverse() with torch.linalg.inv()
• Replaced .transpose(-1, -2) with .mT for matrix transpose
• Conditional torch.jit.optimized_execution for ScriptModule only
• Relaxed ASE dependency from ==3.27.0 to >=3.27.0,<4

Fixed

• AIMNet2-Pd DFT reference corrected from wB97M-D3/CPCM to B97-3c/CPCM (THF)
• Model loading uses weights_only=True by default (security)
• Model download validates HTTP response before writing to disk

Documentation

• Modernized README with pip/uv/conda install instructions
• Updated TorchScript compatibility notes

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Initial release of AIMNet2 MPIP, including model code and pre-trained weights.

Features

• AIMNet2 model architecture and training code
• Core calculator with automatic dense/sparse mode selection
• ASE and PySisyphus integrations
• Long-range Coulomb (Simple, DSF, Ewald) and DFT-D3 dispersion
• Periodic boundary conditions with stress tensor
• GPU acceleration via NVIDIA Warp kernels

Pre-trained Models

• aimnet2 - AIMNet2 wB97M-D3 (default)
• aimnet2_b973c - AIMNet2 B97-3c
• aimnet2_2025 - AIMNet2 B97-3c with improved intermolecular interactions
• aimnet2nse - AIMNet2-NSE for open-shell chemistry
• aimnet2pd - AIMNet2-Pd for palladium-containing systems

Installation

pip install aimnet