SevenNet (Scalable EquiVariance Enabled Neural Network)

SevenNet (Scalable EquiVariance-Enabled Neural Network) is a graph neural network (GNN)-based interatomic potential package. It is designed for parallel molecular dynamics simulations and supports integration with LAMMPS.

Key capabilities include:

• Pretrained Models: Access to pretrained GNN interatomic potentials and a fine-tuning interface.
• ASE Integration: Support for the Python Atomic Simulation Environment (ASE) calculator.
• LAMMPS Support: GPU-parallelized molecular dynamics via LAMMPS integration.
• Dispersion Correction: CUDA-accelerated D3 (van der Waals) dispersion (compatible with LAMMPS pair style d3).
• Advanced Training: Multi-fidelity training for combining databases and fine-tuning with forgetting prevention (using Experience Replay and Elastic Weight Consolidation) for continual learning.

SevenNet-MF-0 is a pretrained model trained on both PBE (+U) and $\mathrm{r}^{2}$ SCAN crystal relaxation trajectories from the Materials Project.

It uses a multi-fidelity training approach with additional 'fidelity-dependent' parameters. These parameters have negligible computational overhead, meaning inference speed is nearly identical to the standard SevenNet-0 model.

Performance (MAE on $\mathrm{r}^{2}$ SCAN):

• Energy: 10.8 meV/atom
• Force: 0.018 eV/Ang.
• Stress: 0.58 kbar

SevenNet provides several pretrained models with varying hyperparameters and training sets, offering different trade-offs between accuracy and speed. These models can be used as interatomic potentials in LAMMPS or loaded via the ASE calculator.

Key model families include:

• SevenNet-Omni: The recommended multi-task model for diverse materials (crystals, molecules, surfaces, etc.).
• SevenNet-Omni-i8 / i12: Higher capacity versions of SevenNet-Omni.
• SevenNet-Nano: A lightweight, distilled version of the mpa task for high-speed simulations.
• SevenNet-MF-ompa: A multi-fidelity model trained on MPtrj, sAlex, and OMat24 datasets.
• SevenNet-omat: A single-task model trained exclusively on OMat24.
• SevenNet-l3i5 / SevenNet-0: Models varying by spherical harmonic degree ($l_{\mathrm{max}}$).

SevenNet-0 (11July2024) is an interatomic potential pre-trained on the MPTrj(CHGNet) dataset. It can be applied directly to any system without additional training. If the accuracy is insufficient for your specific use case, the model can be fine-tuned using a custom dataset.

Key Files in this release:

• checkpoint_sevennet_0.pth: The pre-trained model checkpoint.
• fine_tune.yaml: An example configuration file for fine-tuning the model.
• pre_train.yaml: An example configuration file used during the original pre-training process.

Performance (MAE):

reEWC (rehearsal-enhanced Elastic Weight Consolidation) mitigates catastrophic forgetting when fine-tuning a pretrained model on a new target system. It uses two mechanisms:

1. Experience replay (rehearsal): Replays an old-task "memory" set during each training step.
2. Elastic Weight Consolidation (EWC): Adds a penalty to anchor parameters to their pre-fine-tuning values (theta*), weighted by a precomputed Fisher matrix (F).

Note: reEWC is only supported for single-modal models (e.g., SevenNet-0, SevenNet-Nano). Multi-fidelity models are not currently supported.

When choosing an accelerator, consider the system size and the LAMMPS interface:

1. System Size:

   • Small systems: FlashTP with LAMMPS: Torch generally shows a performance advantage.
   • Large systems (> 10³ atoms): CuEquivariance becomes more efficient as a performance crossover occurs around 1,000 atoms.

2. LAMMPS Interface:

   • FlashTP with LAMMPS: Torch is generally faster than FlashTP with LAMMPS: ML-IAP.
   • CuEquivariance requires LAMMPS: ML-IAP.

To install the core SevenNet package, ensure your environment meets the following requirements:

• Python: >= 3.10
• PyTorch: >= 2.0.0

Once requirements are met, install using pip:

pip install sevenn

You can generate a ready-to-edit configuration file containing both rehearsal and EWC mechanisms using the sevenn preset command.

To create an input file, run:

sevenn preset reewc > input.yaml

Every reEWC key is optional. You can run only one mechanism by removing the other's configuration block:

• To run only rehearsal: Remove the train.continue.ewc keys.
• To run only EWC: Remove the data: replay block.

Once configured, run training as usual:

sevenn train input.yaml -s

TorchSim is an optional GPU-native atomistic simulation engine built on PyTorch that provides batched MD and relaxation. It requires Python >= 3.12. Install it as an extra for SevenNet using:

pip install sevenn[torchsim]

The SevenNet-0 model is an early pretrained model trained on the MPtrj dataset. While newer models are generally more accurate, SevenNet-0 is optimized for the shortest inference time. It is loaded as the default pretrained model when using the ASE (Atomic Simulation Environment) calculator.

To specify this model, use the following keywords:

• 7net-0
• SevenNet-0
• 7net-0_11Jul2024
• SevenNet-0_11Jul2024

Model keywords: `7net-0` | `SevenNet-0` | `7net-0_11Jul2024` | `SevenNet-0_11Jul2024`