Introduction

Motivation

Atomistic Foundation Models have emerged as powerful tools in molecular and materials science. However, the diverse implementations of these open-source models, with their varying architectures and interfaces, create significant barriers for customized fine-tuning and downstream applications.

MatterTune is a comprehensive platform that addresses these challenges through systematic yet general abstraction of Atomistic Foundation Model architectures. By adopting a modular design philosophy, MatterTune provides flexible and concise user interfaces that enable intuitive and efficient fine-tuning workflows.

Key Features

Pre-trained Model Support

Seamlessly work with multiple state-of-the-art pre-trained models including:

  • JMP

  • EquiformerV2

  • M3GNet

  • ORB

Flexible Property Predictions

Support for various molecular and materials properties:

  • Energy prediction

  • Force prediction (both conservative and non-conservative)

  • Stress tensor prediction

  • Custom system-level property predictions

Data Processing

Built-in support for multiple data formats:

  • XYZ files

  • ASE databases

  • Materials Project database

  • Matbench datasets

  • Custom datasets

Training Features

  • Automated train/validation splitting

  • Multiple loss functions (MAE, MSE, Huber, L2-MAE)

  • Property normalization and scaling

  • Early stopping and model checkpointing

  • Comprehensive logging with WandB, TensorBoard, and CSV support