Generative Structure Search accelerates molecular and crystal discovery
A new framework called generative structure search (GSS) has been developed by researchers, merging diffusion-based generation with random structure search (RSS) to effectively predict both stable and metastable molecular and crystal structures. GSS treats these methods as extreme cases of a unified sampling process influenced by learned score fields and physical forces, which allows data priors to enhance sampling efficiency while maintaining energy-guided exploration. GSS successfully identifies a variety of metastable structures at a sampling cost over ten times lower than that of RSS, proving effective even for compositions not included in the training dataset. This research tackles a significant drawback of deep generative models, which tend to overlook rare but significant minima due to biases in training data. The study can be found on arXiv with the identifier 2604.27636.
Key facts
- GSS combines diffusion-based generation and random structure search (RSS) in a unified framework.
- The framework uses learned score fields and physical forces to drive sampling.
- GSS achieves more than tenfold lower sampling cost than RSS for broad coverage.
- It recovers diverse metastable structures across molecular and crystalline systems.
- The method remains effective for compositions outside the training distribution.
- Deep generative models often underexplore rare minima due to training data bias.
- The paper is published on arXiv with identifier 2604.27636.
- Predicting stable and metastable structures is central to molecular and materials discovery.
Entities
Institutions
- arXiv