Newly developed high-throughput methods
for property predictions
make the process of materials design faster and more efficient. Density
is an important physical property for energetic compounds to assess
detonation velocity and detonation pressure, but the time cost of
recent density prediction models is still high owing to the time-consuming
processes to calculate molecular descriptors. To improve the screening
efficiency of potential energetic compounds, new methods for density
prediction with more accuracy and less time cost are urgently needed,
and a possible solution is to establish direct mappings between the
molecular structure and density. We propose three machine learning
(ML) models, support vector machine (SVM), random forest (RF), and
Graph neural network (GNN), using molecular topology as the only known
input. The widely applied quantitative structure–property relationship
based on the density functional theory (DFT–QSPR) is adopted
as the benchmark to evaluate the accuracies of the models. All these
four models are trained and tested by using the same data set enclosing
over 2000 reported nitro compounds searched out from the Cambridge
Structural Database. The proportions of compounds with prediction
error less than 5% are evaluated by using the independent test set,
and the values for the models of SVM, RF, DFT–QSPR, and GNN
are 48, 63, 85, and 88%, respectively. The results show that, for
the models of SVM and RF, fingerprint bit vectors alone are not facilitated
to obtain good QSPRs. Mapping between the molecular structure and
density can be well established by using GNN and molecular topology,
and its accuracy is slightly better than that of the time-consuming
DFT–QSPR method. The GNN-based model has higher accuracy and
lower computational resource cost than the widely accepted DFT–QSPR
model, so it is more suitable for high-throughput screening of energetic
compounds.