Predicting Shock Sensitivity from Differential Scanning Calorimetry Data and Molecular Structure: Beyond the Yoshida Correlation

The Yoshida correlation is widely used in the pharmaceutical
and
fine chemical industry to predict explosivity and shock sensitivity
of chemical substances based on the initiation temperature and enthalpy
of differential scanning calorimetry (DSC) exotherms. We investigate
the origins and accuracy of this correlation (and commonly used modifications
thereof) by applying it to a large data set of 383 compounds, which
are relevant to the pharmaceutical industry, and demonstrate that
the initiation temperature and enthalpy variables are not good predictors
for shock sensitivity. By incorporating structural information (for
the 292 compounds where it was available), we used machine learning
to inform and guide a logistic regression technique to develop a shock
sensitivity model which has a higher overall accuracy (63%) and a
higher accuracy for shock-sensitive compounds (97%) compared to the
original Yoshida correlation (52% overall accuracy, 82% accuracy for
shock-sensitive compounds). This logistic regression model includes
both the original Yoshida variables (DSC initiation temperature and
enthalpy) and also incorporates the oxygen balance (OB100) and the number of energetic nitrogen groups in the molecule.