Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1685
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-17T10:42:06.176647Z
status: published
lens: g_space
sector: science
headline: Machine learning prediction of thermal sensitivity for energetic materials based on multi-source feature fusion
dek: Context Thermal sensitivity, typically characterized by the thermal decomposition temperature (T d ), is a key parameter that reflects the thermal stability and safety of energetic materials, providing essential guidance for their engineering applications, storage, and transportation. However, traditional experimental determination methods suffer from long cycles, high costs, and significant hazards, rendering them unsuitable for efficient, large‑scale material screening. To address this bottleneck, this study s…
gain_title: Combining molecular topological features with first-principles calculated features raised test-set accuracy for predicting decomposition temperature, offering a computational alternative to slow, costly, hazardous experiments for screening energetic materials.
problem_title: (none)
trace_subject: (none)
gain_reading: Combining molecular topological features with first-principles calculated features raised test-set accuracy for predicting decomposition temperature, offering a computational alternative to slow, costly, hazardous experiments for screening energetic materials.
gain_evidence: achieving a test set R 2 of 0.799
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers compared three ways of feeding data to machine learning models that estimate the thermal decomposition temperature of energetic materials: molecular topological features, features from first-principles calculations, and a fusion of the two. The fused set gave the best test-set accuracy, and a Random Forest model with SPXY sample partitioning performed best overall.

Thermal decomposition temperature is a key indicator of the safety and stability of energetic materials, yet measuring it experimentally is slow, expensive, and dangerous. A reliable predictive tool could help narrow candidates before lab work. What remains uncertain is how well the approach generalizes beyond the study's dataset, since the best model still carries an error of roughly 30 °C.
limitation: Even the best fused-feature model retains a test-set RMSE of about 31.6 °C and an R2 below 0.8, so its predictions remain approximate rather than precise.
tag: Evidence-backed gain
key_points: Topological features alone in a Random Forest gave an R2 of 0.721, with RMSE of 39.844 °C and MAE of 29.502 °C. | Calculated features alone performed considerably worse, with an R2 of 0.462 and an RMSE of 42.391 °C. | Fused features reached a test-set R2 of 0.799, cutting RMSE to 31.585 °C and MAE to 24.443 °C. | The improvement appeared across SVR, KRR, and AdaBoost as well as Random Forest.
rundown: The study frames thermal sensitivity, characterized by decomposition temperature, as guidance for engineering use, storage, and transportation of energetic materials. It notes that traditional experimental determination suffers from long cycles, high costs, and significant hazards, which makes large-scale screening impractical.

Among the four models tested, Random Forest combined with SPXY sample partitioning delivered the best overall predictive performance. The authors say the consistent gains across different model types confirm a complementarity between the two kinds of features.

On the methods side, the work used Python with RDKit for topological descriptors, crystal structures from the Cambridge Crystallographic Data Centre, and first-principles calculations with CASTEP and DMol3 using the PBE functional with dispersion corrections.
sources:
- peer_reviewed | Journal of Molecular Modeling | https://doi.org/10.1007/s00894-026-06996-x | 2026-10-09
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

b688634a11dab81f9a6ada95cc5f905e58cbf34bd83d4c2e5f3376d1f41b02b3

previous

0000000000000000000000000000000000000000000000000000000000000000