This project contains the code for Development of Disulfide Epoxy Vitrimers via Transfer Learning: Bridging the Gap between Excellent Comprehensive Properties and Dynamic Cross-Linking Networks published in Macromolecules 2025. Paper

AdaQSPR is a transfer learning-based 3D molecular representation learning framework for predicting the comprehensive properties of epoxy resins. This framework can efficiently predict key performance indicators such as glass transition temperature (Tg), initial thermal decomposition temperature (Td5%), tensile strength (TS), relative permittivity (ε), dielectric loss (tanδ), and electrical breakdown strength (Eb).

.
├── unimol_tools_task_specific/  # Core model code directory
│   ├── models/                 # Model definitions
│   ├── utils/                  # Utility functions
│   ├── tasks/                  # Task definitions
│   ├── config/                 # Configuration files
│   ├── data/                   # Data processing
│   ├── train.py                # Training script
│   ├── predict.py              # Prediction script
├── unimol_tools_domain_specific/  # Domain-specific adaptation code directory
├── data/                       # Dataset directory
│   ├── results-candidate/      # Results of prediction of candidate materials data
│   ├── train/                  # Training data
│   │   ├── Tg.xlsx             # Glass transition temperature data
│   │   ├── Td.xlsx             # Thermal decomposition temperature data
│   │   ├── TS.xlsx             # Tensile strength data
│   │   ├── epsilon.xlsx        # Relative permittivity data
│   │   ├── tan_delta.xlsx      # Dielectric loss data
│   │   ├── Eb.xlsx             # Electrical breakdown strength data
│   │   └── organic_molecular_structure-property_dataset.xlsx # Organic molecular structure-property dataset
│   └── candidate/              # Candidate materials data
├── Tg/                         # Tg prediction related files
├── Td/                         # Td prediction related files
├── TS/                         # TS prediction related files
├── e/                          # ε prediction related files
├── tan/                        # tanδ prediction related files
├── Eb/                         # Eb prediction related files
└── paper.txt                   # Research paper

Vitrimer is a type of covalent adaptable network that can change its topology through exchange reactions, offering excellent properties such as reprocessability, repairability, and degradability. However, compared to traditional thermosetting materials, the thermal, mechanical, and electrical properties of vitrimers may be compromised. This research aims to construct a Quantitative Structure-Property Relationship (QSPR) model to guide the design of disulfide epoxy vitrimers with both excellent comprehensive properties and eco-friendly characteristics.

This project contains two datasets:

1. Organic molecular structure-property dataset: Contains property data of over 100,000 organic molecules
2. Epoxy resin structure-property dataset: Contains experimental data of over 1,700 epoxy resin macroscopic properties

The AdaQSPR model is based on transfer learning and domain-specific adaptation design, mainly consisting of two parts:

1. Domain-specific Adaptation: Based on the Uni-Mol framework, pretrained on the organic molecular structure-property dataset
2. Task-specific Adaptation: Fine-tuned on the epoxy resin structure-property dataset to predict six macroscopic properties of epoxy resins

The research team used the AdaQSPR model to screen 216 different dynamic disulfide epoxy vitrimers and ultimately selected 5 promising candidates for experimental verification. The experimental results proved that these 5 disulfide epoxy vitrimers have excellent comprehensive properties, including:

• Glass transition temperature Tg > 146°C
• Electrical breakdown strength Eb > 34.8 kV/mm
• Four of the materials have tensile strength TS ≥ 60 MPa
• Good repairability and degradability

We open-source the weights of the AdaQSPR model for reproducibility.

Link: https://disk.pku.edu.cn/link/AA547E8FBF09114E7981DA845D7DAD5CCB

Expire time: 2035-08-01 15:53

Please download the weights of Uni-Mol from the origin Repo Uni-Mol.

• Step1: Install the UniMol framework

pip install unimol==1.0.0

• Step2: Install the UniMol-Tools

cd unimol_tools

Then, please follow the instructions in the UniMol-Tools README to install the UniMol-Tools.

• Step3(Optional): Domain-specific Training

cd unimol_tools_domain_specific
python train.py

• Step4(Optional): Train the AdaQSPR model

cd Tg # Take Tg as an example
python trainpoly.py

• Step5: Predict the comprehensive properties of epoxy resins

cd Tg # Take Tg as an example
python pred_poly.py

We would like to thank the following projects for their contributions to this work:

• Uni-Mol

If you have any questions, please contact us at: cheng-handsome at stu.xjtu.edu.cn

If our work has been helpful to you, please consider citing it.

@article{zhang2025development,
  title={Development of Disulfide Epoxy Vitrimers via Transfer Learning: Bridging the Gap between Excellent Comprehensive Properties and Dynamic Cross-Linking Networks},
  author={Zhang, Yucheng and Luo, Junyu and Li, Hao and Li, Wenrui and Li, Wendong and Zhang, Ming and Zhang, Guan-Jun},
  journal={Macromolecules},
  year={2025},
  publisher={ACS Publications}
}