Adaptive virtual design-space generation for efficient, multi-objective materials discovery.
| Language | Project page |
|---|---|
| English | Open |
| 中文 | 打开 |
| 日本語 | 開く |
| 한국어 | 열기 |
The site presents the spatial-adaptive active learning workflow and its key experimental results: 177 mV overpotential, 625 h acidic stability, and a 78× stability improvement.
Many materials-discovery workflows optimize several objectives whose evaluation costs differ drastically. Fast measurements can take minutes; stability tests, simulations, or destructive characterization may take hundreds of hours. Evaluating every objective everywhere is inefficient.
VSGenerator introduces DVSNet, a conditional variational autoencoder that generates an adaptive virtual design space from partially labeled data. It helps focus costly optimization only where promising solutions are most likely to be found.
- Optimize the low-cost objective — identify feasible or high-performing candidates using inexpensive measurements.
- Generate an adaptive virtual space — train DVSNet to represent the region consistent with the first-stage target.
- Optimize the high-cost objective — search only within the focused virtual space, reducing unnecessary experiments.
- Conditional variational autoencoder for adaptive design-space generation
- Multi-objective optimization under non-uniform evaluation costs
- Integration with the Bgolearn framework
- Suitable for closed-loop experiments and data-efficient AI-for-science workflows
pip install VSGeneratorThe complete workflow is available in the VSGenerator tutorial notebook, covering model training, adaptive virtual-space construction, and downstream optimization.
Spatial-adaptive active learning identifies ultra-durable and highly active catalysts for acidic oxygen evolution reaction
Science Bulletin · DOI: 10.1016/j.scib.2025.12.021
@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}
}VSGenerator is released under the MIT License.