Domain-Gated Latent Diffusion for the Discovery of Novel Energetic Materials
Draft paper · [Zenodo DOI 10.5281/zenodo.19821953] · BibTeX
12 DFT-confirmed novel CHNO leads. Headline compound L1 (3,4,5-trinitro-1,2-isoxazole) at rho = 2.09 g/cm^3, D = 8.25 km/s, max-Tanimoto 0.27 to all 65,980 training molecules. Reproducible end-to-end on commodity hardware.
DGLD4Energetic/
paper/ short_paper.html + figs/ Self-contained paper source.
dgld/ LIMO + denoiser + score model + sampler (Python package).
scripts/ Top-level entry points: train_*.py, sample.py, plot_fig*.py.
data/ Small data committed; large data fetched via sidecars.
models/ Sidecar pointers to Zenodo-hosted checkpoints.
experiments/ One subfolder per paper experiment (T1 to T4 + DFT + baselines + ablations).
figures/ Committed final PNG/SVG figures.
docs/ Reproducibility guide, design notes, troubleshooting.
git clone https://github.com/ApartsinProjects/DGLD4Energetic.git
cd DGLD4Energetic
pip install -e .
bash scripts/download_assets.sh # fetches large checkpoints + data from Zenodo
python scripts/plot_fig1.py # writes figures/novelty_vs_D.png| Paper claim | Script | Wall time | Hardware |
|---|---|---|---|
| Fig 1: novelty vs detonation velocity | scripts/plot_fig1.py |
1 min | CPU |
| Fig 19: lead cards | scripts/plot_fig19.py |
1 min | CPU |
| Fig 21: DFT dumbbell | scripts/plot_fig21.py |
1 min | CPU |
| Fig 22: baseline forest plot | scripts/plot_fig22.py |
1 min | CPU |
| Fig 23: quadrant scatter | scripts/plot_fig23.py |
1 min | CPU |
| T1: BDE audit on 12 leads | experiments/t1_bde/ |
TBD | Modal A100 |
| T2: density audit | experiments/t2_density/ |
TBD | Modal A100 |
| T3: denoiser seed variance | experiments/t3_seed_variance/ |
TBD | Modal A100 |
| T4: oxatriazole anchor | experiments/t4_oxatriazole/ |
TBD | Modal A100 |
| 12 DFT-confirmed leads + DNTF anchor | experiments/dft_audit/ |
TBD | Modal A100 |
| Pool fusion (m7) | experiments/m7_pool_fusion/ |
TBD | Modal L4 |
| Multi-seed sampling (m6) | experiments/multi_seed_sampling/ |
TBD | Modal L4 |
| CJ validation vs Cantera | experiments/cj_validation/ |
TBD | CPU |
| H50 sensitivity | experiments/h50_sensitivity/ |
TBD | CPU |
| FCD baselines | experiments/baselines_distribution/ |
TBD | CPU |
| LSTM baseline | experiments/baselines_lstm/ |
TBD | Modal L4 |
| REINVENT baseline | experiments/baselines_reinvent/ |
TBD | Modal L4 |
| SELFIES-GA baseline | experiments/baselines_selfies_ga/ |
TBD | Modal L4 |
| AiZynthFinder retrosynthesis | experiments/aizynth_retro/ |
TBD | CPU |
| Gaussian-control ablation | experiments/gaussian_control/ |
TBD | Modal L4 |
| Tier-gate ablation | experiments/tier_gate_ablation/ |
TBD | Modal L4 |
Large files on Zenodo (DOI 10.5281/zenodo.19821953); pointer files in models/ and data/.
Each sidecar is a small text file with the canonical filename, size, SHA-256, and the Zenodo URL
to wget. See models/*.sidecar for the schema.
@article{aperstein2026dgld,
title = {Domain-Gated Latent Diffusion for the Discovery of Novel Energetic Materials},
author = {Aperstein, Yehudit and Apartsin, Alexey},
journal = {Nature Machine Intelligence},
year = {2026},
doi = {TBD},
note = {DOI 10.5281/zenodo.19821953 (companion repository).}
}- Code: Apache-2.0 (see
LICENSE) - Data + paper: CC-BY-4.0 (see
LICENSE-DATA) - Trained model checkpoints (Zenodo): CC-BY-4.0