IBM/materials

IBM /materials – Foundation Models for Materials (FM4M)

What it is

• A collection of open‑source, pre‑trained foundation models that understand different representations of molecules and materials (SMILES strings, SELFIES strings, molecular graphs, 3‑D atom positions, electron‑density grids, etc.).
• Each model is trained on massive chemistry datasets (up to 91 M SMILES, >1 B molecules, >1.5 M 3‑D structures) and can be used either as‑is for feature extraction or fine‑tuned for downstream tasks such as property prediction, quantum‑property regression, or generative design.
• The repo ships a thin wrapper called FM4M‑Kit that lets you call any of the models through a unified Python API or a simple web UI hosted on Hugging Face Spaces.

Key models

Model	Modality	Core architecture	Typical use‑case
SMI‑TED	SMILES strings	Transformer encoder‑decoder (≈289 M parameters)	Quantum property prediction, sequence‑to‑sequence tasks
SMI‑SSED	SMILES strings	Mamba‑based encoder‑decoder (fast inference)	Same as SMI‑TED but with higher speed
SELFIES‑TED	SELFIES strings	BART‑style transformer	Property prediction and auto‑regressive molecule generation
MHG‑GED	Molecular graphs	GNN encoder + Molecular Hypergraph Grammar decoder	Valid‑molecule generation and graph‑based regression
POS‑EGNN	3‑D atomistic graphs	Equivariant geometric tensor network	Machine‑learning potentials for MD simulations
3DGrid‑VQGAN	3‑D electron‑density grids	VQ‑GAN encoder‑decoder	Compress high‑dimensional quantum data into latent vectors
SMILESDFT‑CLIP	Multimodal (SMILES + 3‑D density)	Joint VQ‑GAN + transformer (contrastive)	Learn aligned representations across string and field modalities
MOL‑MOE	Multi‑view (SMILES, SELFIES, graph)	Mixture‑of‑Experts	Fuse complementary views for higher‑accuracy property prediction
STR‑Bamba	String (SMILES/SELFIES)	Hybrid Transformer‑Mamba‑2	Flexible handling of multiple string encodings
TDiMS	Descriptor (SMILES)	Topological distance of intra‑molecular substructures	Lightweight, interpretable features – no pre‑training needed

How to get started

1. Create a Conda environment

   conda create -n fm4m python=3.9.8
   conda activate fm4m
   

2. Install dependencies

   pip install -r requirements.txt
   # Install torch‑scatter matching your CUDA version, e.g.:
   pip install torch-scatter -f https://data.pyg.org/whl/torch-$(python -c "import torch; print(torch.__version__.split('+')[0])")+cu118.html
   

3. Pick a usage style
   • Direct model – go into models/<model_name>/ and follow the README/notebook there.
   • FM4M‑Kit (recommended) – import the wrapper and call a single function:

     import fm4m
     
     # Extract features from SELFIES‑TED
     feats = fm4m.get_representation(model="selfies-ted", data=my_smiles)
     
     # Train a downstream regressor on a single model
     score = fm4m.single_modal(
         model="MHG-GED",
         x_train=X_train, y_train=y_train,
         x_test=X_test,   y_test=y_test,
         downstream_model="DefaultRegressor"
     )
     
     # Combine several modalities
     score = fm4m.multi_modal(
         model_list=["SELFIES-TED", "MHG-GED", "SMI-TED"],
         x_train=X_train, y_train=y_train,
         x_test=X_test,   y_test=y_test,
         downstream_model="DefaultRegressor"
     )
     

   • Web UI – open the Hugging Face Space linked from the repo and use the drag‑and‑drop interface to upload data, select models, and run a quick XGBoost downstream task.

Repository layout

materials/
├─ models/                # One folder per uni‑modal model (smi_ted/, selfies_ted/, …)
│   ├─ fm4m.py            # Core wrapper utilities
│   └─ …
├─ examples/              # Jupyter notebooks showing end‑to‑end pipelines
│   ├─ fm4m_example.ipynb
│   └─ battery_example.ipynb
├─ requirements.txt       # Python dependencies
└─ README.md (this file)

When to use FM4M

• You have chemical or materials data (SMILES, graphs, 3‑D structures) and need high‑quality embeddings without training a model from scratch.
• You want to compare how different molecular representations affect a downstream prediction.
• You need a pre‑trained ML potential (POS‑EGNN) to run molecular dynamics at near‑DFT accuracy.
• You are interested in generative chemistry (SELFIES‑TED, MHG‑GED) that produces syntactically valid molecules.

Community & support

• Open‑source under an IBM‑friendly license; contributions are welcomed via pull requests.
• Models are hosted on Hugging Face with a GUI for quick exploration.
• The team runs regular tutorials and workshops (MRS 2025, AAAI 2025, NeurIPS 2024, etc.) and maintains a working group in the AI Alliance.

References

• The README lists a paper for each model (Nature Communications Chemistry 2025, NeurIPS 2024, etc.) for deeper technical details.

---

In short, IBM/materials provides a ready‑to‑use toolbox of large‑scale, multimodal foundation models that bring state‑of‑the‑art AI to materials science and chemistry research.