A new benchmark for machine learning applied to powder X-ray diffraction

🌟 Introduction

In this repository we show how to use the SIMPOD dataset, how to create similar data from CIF files and how to train and test some machine learning models to predict space group from simmulated powder X-ray diffraction (PXRD) patterns using SIMPOD.

You can download the SIMPOD dataset from Science Data Bank.

We strongly suggest to download only the 'structures' folder if you only want to use the plane PXRD patterns.

📂 Dataset Structure

The general data structure is shown below. For a more comprehensive view check the Turorial/Tutorial.ipynb file.

Data
├── Structures               <- Structural information and PXRD patterns
│   ├── ID0.json                                       
│   │   ├── ID                          
│   │   ├── space_group                 
│   │   ├── alpha                       
│   │   ├── beta                        
│   │   ├── gamma                       
│   │   ├── a                           
│   │   ├── b                           
│   │   ├── c                           
│   │   ├── intensities             
│   │   └── atoms            
│   └── ID1.json
│   │   
│   │   
│
└── Powder Images            <- Created radial images from PXRD patterns
│   ├── ID0.png
│   └── ID1.png
│   │   
│   │   

🔧 Requirements

• torch >= 2.5.1 (Recommended)
• torchvision >= 0.20.1 (Recommended)
• NVIDIA-GPU + CUDA >= 11.8 (Recommended)

To install the packages clone this repository and run the following command.

conda env --create -f SIMPOD_env.yml

Then, install the corresponding version of torch according to your resources.

📈 Data Loading and Use

We provide a tutorial notebook at Turorial/Tutorial.ipynb that provides a full explanation on how to load and visualize the data. Before running the experiments, the Data folder (following the Dataset Structure section) must be downloaded and placed into the repository main folder.

🔮 Data Creation Process

We created the data following the approach explained on the paper. To replicate this process run the following command, using some example CIF files in the Data_Creation/Files folder.

python Data_Creation/Extract_Info.py

In the Data_Creation/Data_Utils.py file, you will find the main functions used to extract information from the CIF files and generate the simulated diffractograms and images. For a detailed mathematical explanation of the image generation process, go to the Methods section of the paper.

🤖 Machine Learning for Space Group Prediction

Here we provide the codes for the space group prediction using the PXRD patterns from SIMPOD, described in the paper.

👀 Computer Vision Models

To train different computer vision models from scratch run the following command.

CUDA_VISIBLE_DEVICES=<GPU_ID> python Space_Group_Prediction/Train.py --model <model_name> --pretrained <True/False> --epochs <epochs> --training_data <1K/100K/All> --batch_size <batch_size> --lr <learning_rate> --gamma <scheduler_gamma> --patience <scheduler_patience>

To test a trained model, run the following command:

CUDA_VISIBLE_DEVICES=<GPU_ID> python Space_Group_Prediction/Test.py --model <model_name> --weights1 <fold1_best_checkpoint.pt> --weights2 <fold2_best_checkpoint.pt>

You can choose one of the following architectures in "model_name":

• alexnet
• resnet
• densenet
• swin
• swinv2

For example, to train a swinv2 pretrained model, run the following command.

CUDA_VISIBLE_DEVICES=<GPU_ID> python Space_Group_Prediction/Train.py --model swinv2 --pretrained True --epochs 25 --training_data 100K --batch_size 6 --lr 4e-06 --gamma 0.9 --patience 4

Then, to test the model, run the following command:

CUDA_VISIBLE_DEVICES=<GPU_ID> python Space_Group_Prediction/Test.py --model swinv2 --weights1 'swinv2_pretrained_True_lr_4e-06_bs_6_epochs_25_gamma_0.9_patience_4_data_100K_fold1_Best.pt' --weights2 'swinv2_pretrained_True_lr_4e-06_bs_6_epochs_25_gamma_0.9_patience_4_data_100K_fold2_Best.pt'

Note that the weights files names include all the hyperparameters of the model and the fold they were trained on.

If you want to replicate the results of the paper download the pretrained models at Drive. Place the models inside the "Space_Group_Prediction/Models/" folder.

💻 AutoML Models

To train and test the classic machine learning models using the PXRD patterns from SIMPOD with the H2O AutoML library, first download the csv files from the CSVs folder in Science Data Bank, place them inside the Space_Group_Prediction/AutoML_Data/ folder and then run the following command.

python Space_Group_Prediction/automl.py

📃 Citation

Rincón, S., González, G., Macías, M.A. et al. A new benchmark for machine learning applied to powder X-ray diffraction. Sci Data 12, 1186 (2025). https://doi.org/10.1038/s41597-025-05534-3

Bibtex format:

@article{10.1038/s41597-025-05534-3, 
year = {2025}, 
title = {{A new benchmark for machine learning applied to powder X-ray diffraction}}, 
author = {Rincón, Sergio and González, Gabriel and Macías, Mario A and Arbeláez, Pablo}, 
journal = {Scientific Data}, 
doi = {10.1038/s41597-025-05534-3}, 
abstract = {{Although crystal parameter prediction from powder X-ray diffraction has recently attracted the interest of the machine learning community, most existing datasets for this task are private and lack structural diversity. Here, we introduce the Simulated Powder X-ray Diffraction Open Database (SIMPOD), a new dataset that is public and structurally varied. This new benchmark includes 467,861 crystal structures from the Crystallography Open Database (COD) and their powder X-ray diffraction patterns. SIMPOD presents simulated one-dimensional powder X-ray diffractograms and derived two-dimensional radial images to facilitate the adoption of computer vision models for this task. We hope SIMPOD contributes to developing models that improve materials analysis from powder X-ray diffraction.}}, 
pages = {1186}, 
number = {1}, 
volume = {12}
}