This is the official repository of

DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning

Published in Nature Computational Science, 2026

Featured in MIT News

Elton Pan†, Soonhyoung Kwon‡, Sulin Liu†, Mingrou Xie‡, Alexander Hoffman†, Yifei Duan†, Thorben Prein§, Killian Sheriff†, Yuriy Roman-Leshkov‡, Manuel Moliner¶, Rafael Gomez-Bombarelli†, Elsa Olivetti†*

† MIT Materials Science & Engineering, ‡ MIT Chemical Engineering, § TUM, ¶ ITQ-UPV

A subset of the results is also reported in NeurIPS AI for Materials (Oral Spotlight) paper, 2024.

• Python version 3.10.4
• CUDA version >= 11.3

To check your CUDA version, run nvcc --version.

Note: If your CUDA version is earlier than 11.3, you will have to change <YOUR_CUDA_VERSION> to an earlier version for the following lines in env/requirements.txt:

--extra-index-url https://download.pytorch.org/whl/cu<YOUR_CUDA_VERSION>
--find-links     https://data.pyg.org/whl/torch-1.12.1+cu<YOUR_CUDA_VERSION>.html

torch==1.12.1+cu<YOUR_CUDA_VERSION>
torchvision==0.13.1+cu<YOUR_CUDA_VERSION>
torchaudio==0.12.1+cu<YOUR_CUDA_VERSION>

All experiments (training, inference and evaluation) are performed on a Rocky Linux machine with a NVIDIA RTX A5000 GPU (24GB RAM).

1. Clone the repo

git clone https://github.com/eltonpan/zeosyn_gen.git

2. Navigate into the repo

cd zeosyn_gen

3. Create environment

conda create -n zsg python=3.10.4

4. Activate environment

conda activate zsg

5. Install dependencies

pip install -r env/requirements.txt

6. Enable conda environment for jupyter notebook

conda install -c anaconda ipykernel
python -m ipykernel install --user --name=zsg

We demonstrate the DiffSyn model for prediction synthesis recipes for UFI zeolite (Fig. 5 of the manuscript).

To download trained model weights model.pt, run:

wget -O runs/diff/system/run1/model.pt https://www.dropbox.com/scl/fi/vmf5ag87vszlikmlsnlg4/model.pt?rlkey=9p1d2ht0qxr32of0xizsmqxat&st=obgh0a2n&dl=1

This saves the model.pt in the folder runs/diff/system/run1/.

To run inference using DiffSyn, run:

python predict.py

Configurations are defined in predict.py. Here, we generate synthesis recipes for the UFI zeolite with C1COCCN2CCOCCOCCN(CCO1)CCOCCOCC2 as the OSDA. This automatically saves the generated routes at predictions/UFI_C1COCCN2CCOCCOCCN(CCO1)CCOCCOCC2/syn_pred.csv.

Note: Generating 1000 synthesis routes takes ~2 min.

In the same directory as the above, results can be visualized using predictions/UFI_C1COCCN2CCOCCOCCN(CCO1)CCOCCOCC2/compare_pred_and_true.ipynb (Fig. 5a and 5e of the manuscript).

The DiffSyn model is trained on the ZeoSyn dataset, consisting of 23,961 zeolite synthesis routes with 233 zeolite topologies and 921 organic structure-directing agents.

To train DiffSyn, run:

python train_diff.py

Training configurations are defined in train_diff.py. For example, the name of the run is defined as "fname": "run1" This will automatically create a run directory in runs/diff/system/run1. If you run into OOM errors, please lower batch_size in the corresponding config files for each task.

Note: Training will takes ~50 hours. We recommend you download the model weights (see instructions under Demo section).

To evaluate the DiffSyn model, run:

python eval.py

This evaluates the trained model via a 2-step process: First, we generate synthesis recipes for every test system and save it at runs/diff/system/run1/syn_pred_agg-cond_scale_0.75-test.csv. This takes ~200 min. Second, we compute metrics, Wasserstein distance and MAE, of the model and save them at runs/diff/system/run1/wsd_zeo_osda.json and runs/diff/system/run1/reg_zeo_osda.json, respectively.

You should expect to see a mean Wasserstein distance of Mean WSD: 0.423 (Fig. 2a of manuscript).

Baseline models can be trained by running the corresponding train_<MODEL_NAME>.py and evaluated by changing the configs in eval.py.

├── cde
├── data
│   ├── 2024-10-02_K222_and_CHA_OSDA_features.csv
│   ├── 241002_k222_etc_mols_osda_priors_0.pkl
│   ├── augmentation.ipynb
│   ├── cbus-to_be_deleted.csv
│   ├── cifs
│   ├── CVAE_EGNN_embeddings_2023-07-13.csv
│   ├── diffusion_trajectory
│   ├── get_bash_command_for_distance_grid.py
│   ├── get_dummy_graph.py
│   ├── get_zeo_graphs.ipynb
│   ├── gpt4_generated_dict.py
│   ├── iza_codes.py
│   ├── metrics.py
│   ├── osda_descriptors.csv
│   ├── osda_enc_emb.csv
│   ├── prec_rec_vs_hp
│   ├── process_zeosyn.ipynb
│   ├── qt
│   ├── scalers
│   ├── smiles2graph.pkl
│   ├── syn_variables.py
│   ├── utils.py
│   ├── zeo2graph.pkl
│   ├── zeo_enc_emb.csv
│   ├── zeolite_amd_distance_matrix.csv
│   ├── zeolite_binding_energy.csv
│   ├── zeolite_descriptors.csv
│   ├── zeolite_descriptors_for_dendro.csv
│   ├── zeolite_graph_distance.csv
│   ├── zeo_osda_sim-syn_cos_sim.csv
│   ├── zeo_osda_sim-syn_cos_sim-zeo_egnn.csv
│   ├── zeo_osda_sim-syn_mmd_dissim.csv
│   ├── ZEOSYN-2.xlsx
│   ├── ZeoSynGen_dataset.pkl
│   └── ZEOSYN.xlsx
├── env
│   ├── cde.yml
│   ├── requirements.txt
│   ├── zeo_diffusion_metrics_eq.yml
│   ├── zeo_diffusion_metrics.yml
│   └── zeo_diffusion.yml
├── eval.py
├── figs
├── get_diffusion_trajectory.py
├── get_metrics_vs_t.py
├── LICENSE
├── models
│   ├── bnn.py
│   ├── cvae.py
│   ├── diffusion.py
│   ├── gan.py
│   ├── nf.py
│   ├── nn.py
├── notebooks
│   ├── 2024-10-01_OSDAs_to_featurize.ipynb
│   ├── compare_cvae_diff.ipynb
│   ├── compare_cvae_v9_v10.ipynb
│   ├── compare_model_metrics.ipynb
│   ├── compare_model_outputs.ipynb
│   ├── CP-CS1_FAU_LTA.ipynb
│   ├── CS1_MTT_C[N+](C)(C)CCCCCCC[N+](C)(C)C.ipynb
│   ├── CS2_MWW_CCCCCCC[N+](C)(C)C.ipynb
│   ├── CS3_BEC_C[N+](C)(C)CCCCCC[N+](C)(C)C.ipynb
│   ├── CS4_ITG_C[N+]1(C)CCC([N+]2(C)CCCC2)CC1.ipynb
│   ├── CS5_IWR.ipynb
│   ├── CS6_ATO.ipynb
│   ├── dendrogram.ipynb
│   ├── dendrogram_learned_embeds.ipynb
│   ├── get_k222_and_CHA_osdas_features.ipynb
│   ├── get_osda_embeddings.ipynb
│   ├── get_zeolite_embeddings.ipynb
│   ├── metrics_vs_across_hp.ipynb
│   ├── OPT-CS1_CHA.ipynb
│   ├── temp_vs_fwd.ipynb
│   ├── tune_diff_hyperparams.ipynb
│   ├── villaescusa.ipynb
│   ├── visualize_pred_amd.ipynb
│   ├── visualize_pred_bnn.ipynb
│   ├── visualize_pred_cvae-eq.ipynb
│   ├── visualize_pred_cvae-gnn.ipynb
│   ├── visualize_pred_cvae.ipynb
│   ├── visualize_pred_diff.ipynb
│   ├── visualize_pred_gan.ipynb
│   ├── visualize_pred_gmm.ipynb
│   ├── visualize_pred_nf.ipynb
│   ├── visualize_pred_nn.ipynb
│   └── visualize_pred_random.ipynb
├── predictions
│   ├── UFI_C1COCCN2CCOCCOCCN(CCO1)CCOCCOCC2
├── predict.py
├── README.md
├── requirements.txt
├── results
├── runs
│   ├── amd
│   ├── bnn
│   ├── cvae
│   ├── cvae-eq
│   ├── cvae-gnn
│   ├── diff
│   ├── gan
│   ├── gmm
│   ├── nf
│   ├── nn
│   └── random
├── splits
│   └── split_dataset.ipynb
├── train_amd.py
├── train_bnn.py
├── train_cvae-eq.py
├── train_cvae-gnn.py
├── train_cvae.py
├── train_diff.py
├── train_gan.py
├── train_gmm.py
├── train_nf.py
└── train_nn.py

Main scripts

• train_X.py: Train X model
• eval.py: Evaluate model(s)
• predict.py: Run synthesis parameter prediction on a specific zeolite-OSDA pair
• data/get_zeo_graphs.ipynb: Get zeolite graphs resulting in data/zeo2graph.pkl
• data/process_zeosyn.ipynb: Preprocess data/ZEOSYN-2.xlsx to give data/ZeoSynGen_dataset.pkl (dataset object)
• data/augmentation.ipynb: Visualize and analyze data/ZeoSynGen_dataset.pkl for augmentation purposes
• get_diffusion_trajectory.py: Generate and save diffusion trajectory
• get_metrics_vs_t.py: Calculate metrics over diffusion trajectory

Visualizations of results in manuscript

• notebooks/visualize_pred_X.ipynb: Visualizations of model X predictions
• notebooks/compare_cvae_v9_v10.ipynb: Qualitative validation of MMD/WSD metrics by comparing CVAE predictions (from models v9 and v10) to grouth truth
• notebooks/tune_diff_hyperparams.ipynb: Investigation of the effect cond_drop_prob and cond_scale on diffusion performance
• notebooks/dendrogram.ipynb: plot dendrogram of zeolites
• notebooks/compare_model_outputs.ipynb: 2D distribution of synthesis parameters visualization
• notebooks/villaescusa.ipynb: H2O/T vs. zeolite framework density plot (Villaescusa's Rule)
• notebooks/temp_vs_fwd.ipynb: Crystallization temperature distributions vs. zeolite framework density plot
• notebooks/2024-10-01_OSDAs_to_featurize.ipynb: Visualize K222 and CHA OSDAs before featurization
• notebooks/get_k222_and_CHA_osdas_features.ipynb: Extract OSDA features from data/241002_k222_etc_mols_osda_priors_0.pkl and saves preprocessed features as data/2024-10-02_K222_and_CHA_OSDA_features.csv
• notebooks/metrics_vs_across_hp.ipynb: Visualize metrics w.r.t. hyperparameters (t, cond_scale, p_uncond)

Data assets

• data/ZEOSYN-2.xlsx: Cleaned ZeoSyn dataset
• data/zeo2graph.pkl: Dict mapping zeolite IZA code to graph
• data/smiles2graph.pkl: Dict mapping OSDA SMILES to graph
• data/zeolite_descriptors.csv: Zeolite physicochemical descriptors
• data/zeolite_descriptors_for_dendro.csv: Zeolite physicochemical descriptors + some extra info (ring sizes, CBUs etc) for dendrogram plotting
• data/zeolite_binding_energy.csv: Zeolite binding energies to literature OSDAs
• data/zeolite_graph_distance.csv: Zeolite graph and SOAP distances to one another
• data/CVAE_EGNN_embeddings_2023-07-13.csv: EGNN embeddings of zeolites obtained from pretraining CVAE-EGNN on synthesis task
• data/cbus-to_be_deleted.csv: Zeolite CBUs
• data/osda_descriptors.csv: OSDA physicochemical descriptors
• data/iza_codes.py: List of IZA codes
• data/zeolite_amd_distance_matrix.csv: AMD distance matrix featurization of zeolite topologies as reported in Schwalbe-Koda et al (2023) https://github.com/dskoda/Zeolites-AMD/blob/main/data/iza_dm.csv
• data/241002_k222_etc_mols_osda_priors_0.pkl: Features of K222 and CHA OSDAs (from Science paper)
• data/2024-10-02_K222_and_CHA_OSDA_features.csv: Preprocessed K222 version of data/241002_k222_etc_mols_osda_priors_0.pkl
• data/get_dummy_graph.py: Get placeholder graphs for amorphous phases and zeolites with no CIF files
• data/get_bash_command_for_distance_grid.py: script to construct bash script for Zeo++
• data/syn_variables.py: Column names of synthesis parameters
• data/utils.py: Helper functions for data preprocessing and visualization

Archive: Precursor generation (for future work)

• prec_rec/prepare_prec_dataset.ipynb: Prepare precursor generation dataset
• prec_rec/precusors_raw.py: Dictionary mapping raw text to elemental identity and common names
• prec_rec/precusors_clean.py: Dictionary mapping raw text to elemental identity and common names (cleaned by Soon 2024-05-08)
• prec_rec/prec_dataset/prec_dataset_X.csv: Specific split for precursor generation dataset
• prec_rec/prec_dataset/prec_dataset_X.csv: Specific split for precursor generation dataset
• visualize_prec_dataset.ipynb: Guide on how to access key parts of precursor generation dataset

Note: The cde environment is required to run training and inference of Gausssian mixture models (gmm).

If you have any questions, please contact us at eltonpan@mit.edu or elsao@mit.edu.

If you use this dataset or code, please cite this paper:

@article{pan2026diffsyn,
  title={DiffSyn: a generative diffusion approach to materials synthesis planning},
  author={Pan, Elton and Kwon, Soonhyoung and Liu, Sulin and Xie, Mingrou and Hoffman, Alexander J and Duan, Yifei and Prein, Thorben and Sheriff, Killian and Roman-Leshkov, Yuriy and Moliner, Manuel and others},
  journal={Nature Computational Science},
  pages={1--13},
  year={2026},
  publisher={Nature Publishing Group US New York}
}

• Test conda installation on Linux systems
• Test conda installation on non-Linux system
• Add link to paper
• Update Bibtex (after issue is out)
• Add Colab notebook option