Repository Description

This repository contains the model codes for the paper:
​​《d-band center-guided high-fidelity generative model for inverse materials design》

A crystal diffusion generative model that supports d-band center and space group information as guiding conditions. Several novel strong adsorption materials were screened and demonstrated.

If the paper is useful to you, kindly consider citing it.

Subsequent Optimization

In the original code, the node feature in the ​denoiser​ underwent two rounds of message passing. We found that this weakened the feature strength of the d-band center. Therefore, we optimized the model by retaining only a single round of message passing. After retraining the model, t-SNE visualizations reveal that the denoiser’s capability to capture features of the d-band center is enhanced.
Therefore, we recommend using the ​optimized verson​. <img width="1734" height="668" alt="拼图" src="https://github.com/user-attachments/assets/b3d55e65-9e2a-431d-9bd7-6ec49c64dac3" />

Overview

Model Training and Inference

Create a Conda Environment and Install Dependencies

conda create -n dbanddiff python=3.8 -y
conda activate dbanddiff
pip install -r requirements.txt

Training

Step 1: Unzip ​optimized verson.zip​

Step 2: Run the Training.ipynb

conda activate dbanddiff
jupyter notebook Training.ipynb    

Inference

Step 1: Download the model weights from Google Drive and place it in the same directory as "Generation.ipynb":

https://drive.google.com/file/d/1p4ut00OeAecBmBc56wCsBk_2V2AVaga6/view?usp=sharing

Step 2: Run the Generation.ipynb

conda activate dbanddiff    
jupyter notebook Generation.ipynb

Dependency

Including ​PyTorch Geometric​​, ​PyTorch​​, ​Pymatgen​, etc. Please refer to the ​requirement.txt​​ for details.

Support

Thank ​Rui Jiao (Tsinghua University)​ for providing detailed guidance on the details of the source framework ​DiffCSP++​. For any questions, please raise issues or contact wogaho1999@gmail.com.