Open Source Assets to support AI for Science projects

Tip: Use the search box at the top of this page to find specific content.

Foundation Model for Materials (FM4M)

Foundation Model for Materials (FM4M) is a suite of models pre-trained on billions of multimodal molecular data built on SMILES, SELFIES, and molecular graphs. You can bring your own data to address general and domain-specific prediction tasks (e.g. battery materials). Available at:

Github: https://github.com/IBM/materials/

HuggingFace: https://huggingface.co/collections/ibm-research/materials-6798751004389cf7195eae6a

Typical use cases are property prediction, structure generation, etc. IT-skilled users can leverage Jupyter Notebooks. Chemistry experts can use a GUI-based web application, or a LLM-based chat interface.

Biomedical Foundation Models (BMFM)

Biomedical Foundation Models (BMFM) leverage multi-modal data of different types, including drug-like small molecules and proteins (covering a total of more than a billion molecules), single-cell RNA gene-expression levels, whole genome DNA sequence, and other biomedical data. These can be used for various purposes such as antibody property prediction, gene perturbation prediction, cell type annotation, single-nucleotide polymorphism/disease association, trial design, etc. More information at:

Blog: https://research.ibm.com/projects/biomedical-foundation-models

Github: https://github.com/BiomedSciAI

HuggingFace: https://huggingface.co/collections/ibm/biomed-671fc7694b2e5a664a9f098e

Prithvi-IBM-NASA Foundation Models for Earth

IBM and NASA have teamed up to create a family of AI foundation models for Earth called Prithvi. Datasets for pretraining consisted of NASA’s Harmonized Landsat and Sentinel-2 satellite data and NASA’s MERRA-2 climate data. You can do fine-tuning using your own data to build general and domain-specific models (e.g., lat/lon polygons with crop type information). These geospatial foundation models make AI accessible for open science users, startups, and enterprises by simplifying model training and deployment. Available at:

Github: https://github.com/NASA-IMPACT/

HuggingFace: https://huggingface.co/ibm-nasa-geospatial

Depending on the skill level of the users, they can use Jupyter Notebooks or Terratorch and GEO-Bench-2

Typical use cases are Earth Observations (e.g. flood detection, wildfire scars detection, crop health, etc.) and Weather & Climate (e.g. climate model downscaling, hurricane track & intensity prediction, weather forecasting, etc.) These models deliver high accuracy and reliability while excelling in tasks with limited labelled data.