README.md

metadata

license: mit
datasets:
  - chaitjo/QM9_ADiT
  - chaitjo/MP20_ADiT
  - chaitjo/QMOF150_ADiT
  - chaitjo/GEOM-DRUGS_ADiT
tags:
  - chemistry
  - materials
  - molecules
  - crystals
  - diffusion
  - transformer
  - latent-diffusion
  - all-atom
model-index:
  - name: ADiT
    results:
      - task:
          type: unconditional-molecule-generation
        dataset:
          name: QM9
          type: QM9
        metrics:
          - name: Validity Rate
            type: Validity Rate
            value: 94.45
        source:
          name: Unconditional Molecule Generation
          url: >-
            https://paperswithcode.com/sota/unconditional-molecule-generation-on-qm9
      - task:
          type: unconditional-crystal-generation
        dataset:
          name: MP20
          type: MP20
        metrics:
          - name: Validity Rate
            type: Validity Rate
            value: 91.92
          - name: DFT S.U.N. Rate
            type: DFT S.U.N. Rate
            value: 6
        source:
          name: Unconditional Crystal Generation
          url: >-
            https://paperswithcode.com/sota/unconditional-crystal-generation-on-mp20
      - task:
          type: unconditional-molecule-generation
        dataset:
          name: GEOM-DRUGS
          type: GEOM-DRUGS
        metrics:
          - name: Validity Rate
            type: Validity Rate
            value: 95.3
        source:
          name: Unconditional Molecule Generation
          url: >-
            https://paperswithcode.com/sota/unconditional-molecule-generation-on-geom
library_name: transformers

All-atom Diffusion Transformers

Independent reproduction of the paper "All-atom Diffusion Transformers: Unified generative modelling of molecules and materials", by Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram*, and Zachary W. Ulissi* from FAIR Chemistry at Meta, published at ICML 2025 (* Joint last author).

All-atom Diffusion Transformers (ADiTs) jointly generate both periodic materials and non-periodic molecular systems using a unified latent diffusion framework:

• An autoencoder maps a unified, all-atom representations of molecules and materials to a shared latent embedding space; and
• A diffusion model is trained to generate new latent embeddings that the autoencoder can decode to sample new molecules or materials.

Note that these checkpoints are the result of an independent reproduction of this research by Chaitanya K. Joshi, and may not correspond to the exact models/performance metrics reported in the final manuscript.

These checkpoints can be used to run inference as described in the README on GitHub.

Here is a minimal notebook for loading an ADiT checkpoint and sampling some crystals or molecules:

Examples of 10,000 sampled crystals and molecules are also available on HuggingFace:

• Crystals as CIF files (MP20)
• Molecules as PDB files (QM9)
• Molecules as PDB files (GEOM-DRUGS)

Citation

Accepted as a conference paper at ICML 2025. Also presented as a Spotlight talk at ICLR 2025 AI for Accelerated Materials Design Workshop. ArXiv link: All-atom Diffusion Transformers: Unified generative modelling of molecules and materials

@inproceedings{joshi2025allatom,
  title={All-atom Diffusion Transformers: Unified generative modelling of molecules and materials},
  author={Chaitanya K. Joshi and Xiang Fu and Yi-Lun Liao and Vahe Gharakhanyan and Benjamin Kurt Miller and Anuroop Sriram and Zachary W. Ulissi},
  booktitle={International Conference on Machine Learning},
  year={2025},
}