IBM Materials is a foundation model for materials science that provides pre-trained uni-modal and multi-modal models for predicting and generating molecular properties and structures.
The tool addresses the challenge of applying machine learning to materials discovery and property prediction by offering a modular architecture with separate uni-modal models trained on distinct molecular representations including SMILES, SELFIES, molecular graphs, 3D atom positions, and 3D density grids. Users can access individual models directly or combine them through late-fusion algorithms to create more powerful multi-modal representations. The project provides fm4m-kit, a wrapper that simplifies access to all model capabilities through straightforward methods, and makes models available on Hugging Face with a graphical interface.
The tool suits researchers and practitioners working on downstream materials science tasks who want to leverage pre-trained representations rather than training from scratch. It is designed for fine-tuning with domain-specific data to support both predictive tasks, such as property estimation, and generative tasks. The modular architecture allows users to start with individual uni-modal models and progressively incorporate additional representations as needed, making it flexible for projects of varying scope and complexity.
The project typically responds to issues and pull requests within one to two weeks. Work in the issue tracker centers on enhancement requests and questions from users, suggesting an active engagement with the community around extending capabilities and clarifying usage.
