AIMNet2: Machine-Learned Interatomic Potential¶

What is AIMNet2?¶

AIMNet2 is a neural network potential for fast and accurate atomistic simulations. Built on PyTorch, it provides:

Accurate predictions for neutral, charged, organic, and elemental-organic systems
Fast inference on both CPU and GPU
Integration with popular simulation packages (ASE, PySisyphus)
Configurable long-range electrostatics (DSF, Ewald) for periodic systems

AIMNet2 combines a graph neural network architecture with flexible long-range interactions, making it suitable for molecular dynamics, geometry optimization, and property prediction across diverse chemical systems.

Explore AIMNet2¶

Choose a Model¶

Find the right model for your chemistry -- general organic, open-shell radicals, Pd catalysis, non-covalent interactions, or high-throughput screening.

Model Selection Guide

Learn the Basics¶

Step-by-step tutorials from your first single-point calculation through geometry optimization, molecular dynamics, and batch processing.

Start with Single-Point Calculations

Advanced Workflows¶

Conformer search, reaction paths and transition states, transition metal catalysis, non-covalent interactions, and charged systems.

Advanced Use Cases

Quick Start¶

from aimnet.calculators import AIMNet2Calculator

# Load model
calc = AIMNet2Calculator("aimnet2")

# Run inference
result = calc({
    "coord": coords,    # (N, 3) array in Angstrom
    "numbers": numbers, # (N,) atomic numbers
    "charge": 0.0,      # molecular charge
}, forces=True)

# Access results
energy = result["energy"]   # eV
forces = result["forces"]   # eV/Angstrom
charges = result["charges"] # partial charges

Available Models¶

AIMNet2 provides five model families covering a wide range of chemistry -- from general organic molecules to open-shell radicals and palladium catalysis. Each model has ensemble members (append _0 to _3) for uncertainty estimation.

See the Model Selection Guide for a detailed comparison and decision flowchart.

Installation¶

Basic installation:

pip install aimnet

With optional features:

# ASE integration
pip install "aimnet[ase]"

# PySisyphus integration
pip install "aimnet[pysis]"

# Training tools
pip install "aimnet[train]"

Requirements: Python 3.11 or 3.12. GPU support optional (PyTorch with CUDA).

Documentation Guide¶

First Steps¶

Getting Started - Installation, setup, and verifying your environment

Models¶

Model Selection Guide - Decision flowchart for choosing the right model
Architecture Overview - AEV descriptors, ConvSV, charge equilibration internals
AIMNet2 (wB97M-D3) - General-purpose default model
AIMNet2-2025 - Recommended B97-3c model (supersedes B97-3c)
AIMNet2-B97-3c - Legacy B97-3c (for reproducibility)
AIMNet2-NSE - Open-shell and radical chemistry
AIMNet2-Pd - Palladium-containing systems

Tutorials¶

Single-Point Calculations - Your first calculation
Geometry Optimization - Structure relaxation with ASE
Molecular Dynamics - NVT/NPT simulations
Periodic Systems - Crystals, surfaces, and PBC
Batch Processing - Processing molecular datasets
Performance Tuning - compile_model=True, GPU optimization

Advanced Use Cases¶

Open-Shell Chemistry - Radicals and spin states
Pd Catalysis - Transition metal reactions
Non-Covalent Interactions - H-bonding, pi-stacking
Conformer Search - Conformational sampling
Reaction Paths & TS - Transition states with PySisyphus
Charged Systems - Ions and zwitterions

Reference¶

Calculator API - Comprehensive reference for AIMNet2Calculator
Long Range Methods - Coulomb and dispersion methods
Model Format - Understanding model files and metadata
Training - Training custom models
CLI Reference - Command-line tools

API Reference¶

Calculators - AIMNet2Calculator, AIMNet2ASE, AIMNet2Pysis
Modules - Core neural network modules
Data - Dataset and sampling utilities
Config - Configuration utilities

Support and Contributing¶

Issues: GitHub Issues
Discussions: GitHub Discussions
Repository: github.com/isayevlab/aimnetcentral

Citation¶

If you use AIMNet2, please cite:

@article{aimnet2,
  title={AIMNet2: A Neural Network Potential to Meet Your Neutral, Charged, Organic, and Elemental-Organic Needs},
  author={Anstine, Dylan M and Zubatyuk, Roman and Isayev, Olexandr},
  journal={Chemical Science},
  volume={16},
  pages={10228--10244},
  year={2025},
  doi={10.1039/D4SC08572H}
}

License¶

See LICENSE for details.