All-atom Diffusion Transformers: Unified generative modelling of molecules and materials
Authors: Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, Zachary Ward Ulissi
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on MP20, QM9 and GEOMDRUGS datasets demonstrate that jointly trained ADi T generates realistic and valid molecules as well as materials, obtaining state-of-the-art results on par with molecule and crystal-specific models. |
| Researcher Affiliation | Industry | 1Fundamental AI Research (FAIR) at Meta 2University of Cambridge 3MIT. |
| Pseudocode | Yes | Algorithm 1: Pseudocode for VAE encoder E |
| Open Source Code | Yes | Open source code: https://github.com/facebookresearch/allatom-diffusion-transformer |
| Open Datasets | Yes | For our main experiments, we train models on periodic crystals from MP20 and non-periodic molecules from QM9. MP20 (Xie et al., 2022) contains 45,231 metastable crystal structures from the Materials Project (Jain et al., 2013)... QM9 (Wu et al., 2018) consists of 130,000 stable small organic molecules... |
| Dataset Splits | Yes | We split the data following prior work (Xie et al., 2022; Hoogeboom et al., 2022) to ensure fair comparisons. |
| Hardware Specification | Yes | Both models are trained to convergence for at most 5000 epochs up to 3 days on 8 V100 GPUs. |
| Software Dependencies | No | The paper mentions software like RDKit, Py Mat Gen, CHGnet, and MOFChecker but does not provide specific version numbers for any of them. For example, it mentions 'constructing the molecule via RDKit' and 'build the crystal structure using Py Mat Gen' but without specific version numbers. |
| Experiment Setup | Yes | We sequentially train the first-stage VAE and then the second-stage Di T using Adam W optimizer with a constant learning rate 1e 4, no weight decay, and batch size of 256. We use exponential moving average (EMA) of Di T weights over training with a decay of 0.9999. |