AI2Pot v1.2.2

Main updates:

1. Fix bug introduced in AI2Pot v1.2.1: Can't export .pt file for MTP model.

AI2Pot v1.2.1

Main updates:

1. Added atom-count normalization for energy and virial losses for training.

2. Refactored the linearMtpToEsitesOp operator.

3. Added linearMtpToEsitesJacobianOp for computing the Jacobian of Linear MTP site energies with respect to model parameters.

AI2Pot v1.2.0

Summary

• Add typewise ZBL for MTP, NEP and NNMTP in PR #19
• Supports one line command installation: pip install <wheel_file> in PR #20 (containing detailed steps). If you want to use it, you should meet:
  • Python 3.11 or 3.12;
  • CUDA 12.4 on Linux and CPU builds on Linux or macOS;
  • torch 2.5.0

AI2Pot v1.1.3

Summary

Modify default f_wgt_start and v_wgt_end to improve accuracy in PR #17

AI2Pot v1.1.2

Summary

This release improves the training stability of NNMTP and NEP in AI2Pot.

Key updates include:

• Add configurable weight_decay support to LitPotentialBase in PR#16
• Switch the optimizer to AdamW for decoupled weight decay.
• Improve the numerical stability of the tanh implementation to prevent overflow for large hidden values in PR#16

AI2Pot v1.1.1

Summary

• Unified AI2Pot and LAMMPS neighbor-list indexing in PR #8
• Added GitHub Actions workflow for automated wheel builds in PR #9
• Update default force and virial weight settings during training in PR #12
• Improved code consistency and release infrastructure

AI2Pot v1.1.0

• Added support for the NNMTP (Neural Network Moment Tensor Potential) model for training and inference.
• Updated related C++/CUDA bindings and backend libraries to support independent CPU/GPU loading.

AI2Pot v1.0.0

• Initial release of AI2Pot
• Support MTP / NEP training and inference
• C++/CUDA kernel optimization
• LAMMPS and ASE interface added