AI Discovers New Energetic Material via Domain-Gated Diffusion
A new generative AI model, Domain-Gated Latent Diffusion (DGLD), has been developed by researchers, leading to the identification of 12 innovative energetic materials. Among these is 3,4,5-trinitro-1,2-isoxazole (L1), which boasts a calculated density of 2.09 g/cm³ and a detonation velocity of 8.25 km/s. Notably, no new HMX-class compounds have been reported in the last fifteen years. The model tackles the sparse-label issue, as only about 3,000 of the roughly 66,000 labeled CHNO molecules have reliable experimental or DFT-quality data. DGLD employs a label-quality gate during training, multi-task score-model guidance for sampling, and a four-stage chemistry-validation process culminating in a first-principles DFT audit. The compound L1 differs structurally from all 65,980 training molecules. This research was published on arXiv (2605.26540).
Key facts
- DGLD discovered 12 DFT-confirmed novel energetic materials.
- Headline compound L1 reaches ρ_cal=2.09 g/cm³ and D_K-J,cal=8.25 km/s.
- No new HMX-class compound has been disclosed in fifteen years.
- Only ~3k of ~66k labeled CHNO molecules have experimental or DFT-quality measurements.
- DGLD uses a label-quality gate, multi-task score-model guidance, and a four-stage validation funnel.
- L1 is structurally dissimilar from all 65,980 training molecules.
- The research was published on arXiv with ID 2605.26540.
- The model is designed to avoid memorization or uncalibrated extrapolation.
Entities
Institutions
- arXiv