CederGroupHub/chgnet

What it solves

CHGNet is a pretrained universal neural network potential designed for atomistic modeling of materials. It solves the problem of computationally expensive Density Functional Theory (DFT) calculations by providing a fast, near-DFT accuracy alternative for predicting energy, forces, stress, and magnetic moments of crystal structures across the periodic table.

How it works

It is a Crystal Hamiltonian Graph Neural Network (CHGNet) pretrained on the Materials Project trajectory (MPtrj) dataset, which includes over 1.5 million structures from 146k compounds. The model regularizes atom features with DFT magnetic moments to better capture electron interactions and charge distribution, allowing it to perform charge-informed modeling.

Who it’s for

Researchers in materials science and computational chemistry who need to perform fast structure optimization, molecular dynamics (MD) simulations, and materials stability predictions without the high cost of traditional DFT calculations.

Highlights

• Universal Potential: Pretrained on a massive dataset spanning the whole periodic table.
• Multi-property Prediction: Predicts energy, forces, stress, and magnetic moments (magmom).
• Integration: Works with ASE (Atomic Simulation Environment) and LAMMPS for molecular dynamics.
• Fine-tuning: Supports fine-tuning on custom datasets to increase precision for specific systems of interest.
• Hardware Support: Runs on CPU, NVIDIA GPUs (CUDA), Intel GPUs (XPU), and Apple silicon (MPS).