DGLD: Domain-Gated Latent Diffusion for the Discovery of Novel Energetic Materials

Researchers introduce Domain-Gated Latent Diffusion (DGLD), an AI method that discovered 12 novel energetic materials using generative diffusion models with quality-gated training and multi-task guidance. The breakthrough identified two lead compounds with performance metrics rivaling HMX-class materials for the first time in 15 years, validated through DFT simulations and released with open-source code.

DGLD addresses a critical materials science challenge: discovering high-performance energetic compounds in an extremely sparse-label regime where only ~3,000 of 66,000 known CHNO molecules have reliable experimental or computational validation. The researchers tackled the fundamental problem that naive generative models either memorize outliers or fail to generalize meaningfully. Their solution combines a label-quality gate during training to prioritize high-confidence data, multi-task score-model guidance at inference to balance novelty and performance, and a four-stage validation funnel culminating in first-principles DFT audits.

The methodology represents a significant advance in AI-assisted materials discovery because it solves the exploration-exploitation tension: DGLD simultaneously achieved chemical novelty (compounds structurally dissimilar from all 65,980 training molecules) while landing in the "productive quadrant" of desired performance metrics. The lead compound L1 achieves a detonation velocity of 8.25 km/s, while co-lead E1 reaches 9.00 km/s—metrics competitive with HMX-class standards. Critically, DGLD outperformed comparable baselines: SMILES-LSTM memorized 18.3% of outputs, SELFIES-GA lost performance under DFT validation, and REINVENT 4 plateaued at suboptimal velocities.

This work signals maturation in generative AI for materials science. The release of code, checkpoints, and 918 mined hard negatives democratizes access to these techniques, enabling rapid iteration cycles measured in GPU-days rather than laboratory months. For defense and civilian applications requiring propellant efficiency, the 15-year innovation gap in this domain makes this discovery consequential.

• →DGLD discovered two novel energetic compounds with HMX-competitive performance metrics after 15 years without breakthroughs in this materials class.
• →The method combines label-quality gating and multi-task guidance to solve sparse-label generative modeling, avoiding both memorization and poor generalization.
• →All 12 discovered leads underwent DFT-level validation; only DGLD consistently landed in the productive quadrant of novelty and target performance.
• →Open-source release of code and datasets accelerates reproducibility and enables rapid validation cycles at minimal computational cost.
• →Success demonstrates AI's capacity to navigate high-dimensional chemical space where traditional discovery methods stalled, with direct applications to propulsion and defense.