For any inquiries or assistance, feel free to contact Mr. CAO Bin at: 📧 Email: bcao686@connect.hkust-gz.edu.cn Cao Bin is a PhD candidate at the Hong Kong University of Science and Technology (Guangzhou), under the supervision of Professor Zhang Tong-Yi. His research focuses on AI for science, especially intelligent crystal-structure analysis and discovery. Learn more about his work on his homepage.

---

Dynamic Virtual Space Generation Neural Network

Paper ：Science Bulletin

Scientific Motivation

In data-driven materials discovery and other AI-for-science applications, multi-objective optimization problems often involve strongly imbalanced evaluation costs. Some target properties can be measured or labeled at low cost, while others require expensive experiments, long simulations, or destructive characterization.

Conventional Bayesian optimization or active-learning strategies typically assume uniform evaluation costs and therefore suffer from severe inefficiency when applied to such asymmetric settings. This limitation becomes a critical bottleneck in closed-loop experimental design.

Method Overview

VSGenerator introduces DVSNet, a conditional variational autoencoder designed to learn adaptive virtual design spaces from partially labeled data.

The core idea is to decouple multi-objective optimization into sequential stages:

1. Low-cost objective optimization The objective that is inexpensive to evaluate is optimized first, producing a set of feasible or high-performing candidates.

2. Adaptive virtual space construction DVSNet is trained to learn a conditional latent representation that captures the design subspace consistent with the constraints imposed by the first objective.

3. Constrained high-cost optimization The expensive objective is optimized exclusively within the generated virtual space, thereby avoiding unnecessary labeling for each sample.

This strategy enables cost-aware exploration, substantially reducing the number of high-cost evaluations required to identify optimal solutions.

Key Features

• Conditional variational autoencoder for adaptive design-space generation
• Explicit support for multi-objective optimization with non-uniform evaluation costs
• Seamless integration with the Bgolearn framework
• Suitable for closed-loop experimental design and data-efficient materials discovery

Installation

Install the package via pip:

pip install VSGenerator

Usage and Reproducibility

A complete, reproducible workflow is provided in the tutorial notebook:

https://github.com/Bgolearn/VSGenerator/blob/main/tutorial/VSGenerator_Tutorial.ipynb

The tutorial demonstrates:

• Model initialization and training of DVSNet
• Construction of adaptive virtual spaces
• Integration with downstream optimization pipelines

License

This project is distributed under the MIT License. See the LICENSE file for licensing details.

---

Citation

If you use VSGenerator or DVSNet in your research, please cite the associated paper:

@article{Cao2025SpatialAdaptiveAL,
  title   = {Spatial-adaptive active learning identifies ultra-durable and highly active catalysts for acidic oxygen evolution reaction},
  author  = {Cao, Bin and
             Qin, Yin and
             Luo, Yan and
             Ying, Zhehan and
             Yan, Zilin and
             Weng, Lu-Tao and
             Li, Kaikai and
             Zhang, Tong-Yi},
  journal = {Science Bulletin},
  year    = {2025},
  doi     = {10.1016/j.scib.2025.12.021},
  url     = {https://www.sciencedirect.com/science/article/pii/S2095927325012678?via%3Dihub#m0005}
}