A label-free domain-knowledge-informed framework for transferability evaluation of synthetic accessibility scoring in energetic chemical space
Synthetic accessibility (SA) evaluation serves as a critical bridge between theoretical material design and experimental realization, and SA scoring has gained significant traction in the data-driven paradigm. However, the scarcity of high-quality annotated synthesis difficulty data poses a formidable challenge for cross-domain assessment of existing scoring methods, particularly when transferring models from the pharmaceutical sector to energetic materials. To address this, we propose a domain-knowledge-driven evaluation framework tailored for unlabeled data. A multi-stage screening protocol integrating structural constraints, property calculation, and statistical monotonicity was implemented to construct a statistically controlled dataset of 28,195 candidate molecules. Unsupervised learning was then employed to characterize the distribution patterns of eight SA scoring models within the energetic chemical space, enabling preliminary transferability assessment without labels. Domain-knowledge-based validation, including boundary sample analysis and energy descriptor evaluation, confirmed that SAscore and BR-SAScore exhibit the most consistently favorable transferability. Using RDX as a benchmark, we delineated a high-energy, easy-to-synthesize chemical space, revealing that heteroatom-rich linkages with moderate scaffold complexity are pivotal for novel high-energy-density materials. This work establishes a label-free framework for evaluating SA model transferability in neutral nitro-containing CHON chemical space, providing a robust methodological reference for developing energetic-material-specific scoring models.
Authors
- Linyuan Wen (ORCID: https://orcid.org/0000-0002-2603-0078)
- Chao Chen (ORCID: https://orcid.org/0000-0001-9743-3360)
- Tao Yu (ORCID: https://orcid.org/0009-0001-8083-2418)
- Zihua Li
- Yingzhe Liu (ORCID: https://orcid.org/0000-0003-4150-1111)
- Zhixiang Zhang
- Yining Zhang (ORCID: https://orcid.org/0009-0004-6004-2939)
Publication Details
- Journal
- npj Computational Materials
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1038/s41524-026-02396-0
- Primary Topic
- Energetic Materials and Combustion
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Natural Science Foundation of Shaanxi Province