Abstract
In order to predict the impact sensitivity of high explosives, we designed and evaluated several models based on the trigger linkage hypothesis and the Arrhenius equation. To this effect, we calculated the heat of detonation, temperature of detonation, and bond dissociation energy for 70 energetic molecules. The bond dissociation energy divided by the temperature of detonation proved to be a good predictor of the impact sensitivity of nitroaromatics, with a coefficient of determination (R2) of 0.81. A separate Bayesian analysis gave similar results, taking model complexity into account. For nitramines, there was no relationship between the impact sensitivity and the bond dissociation energy. None of the models studied gave good predictions for the impact sensitivity of liquid nitrate esters. For solid nitrate esters, the bond dissociation energy divided by the temperature of detonation showed promising results (R2 = 0.85), but since this regression was based on only a few data points, it was discredited when model complexity was accounted for by our Bayesian analysis. Since the temperature of detonation correlated with the impact sensitivity for nitroaromatics, nitramines, and nitrate esters, we consider it to be one of the leading predictive factors of impact sensitivity for energetic materials.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
Introduction
By using quantum chemical and thermodynamic calculations, new energetic molecules can be designed and characterized in terms of their geometry, density, and performance as explosives and propellants [1,2,3,4,5]. It is important to predict the sensitivity of energetic materials, but despite considerable efforts made during the last decades, developing a reliable and general method is still challenging [6,7,8,9,10,11]. By the sensitivity of an energetic material, we refer to its susceptibility to initiate due to external thermal, mechanical, or electrostatic stimuli. The study of the underlying causes that govern sensitivity is not only important for understanding liquid and solid-state phenomena in general but above all for ensuring safe handling, transport, and storage of energetic materials.
One of the most well-known measures of the sensitivity of an explosive is its impact sensitivity, which is determined by dropping a mass upon the sample, measuring the critical point at which a pre-decided fixed percentage of the drops will lead to an explosion. We refer to this critical point as the critical impact level of the material, which may either be given as the critical height (cm) or potential energy (J). Hence, the impact sensitivity and the critical impact level are inversely correlated.
The impact sensitivity is related to macroscopic parameters such as particle size, crystals defects, polymorphism, and crystal orientation. Defects play a particularly prominent role since they form hot-spots under fast compression of the material. The initiation process can be divided into two steps. First, the material is compressed and deformed, leading to heating of the hot-spots. In the second step, the material inside and surrounding the hot-spots self-ignites and propagates into an explosion, provided that the hot-spot temperatures are sufficiently high. The critical temperature at which the explosive self-ignites during impact has been measured to be between 390 and 1060∘ C [12, 13].
There are many operational factors which also affect the critical impact level measured, such as the type of fallhammer used, the test procedure itself, and operator-related judgment of explosion/no-explosion. In addition, the impact sensitivity will depend on the thickness and size of the sample [14, 15]. In general, the measured critical impact level depends on a variety of experimental factors in addition to molecular-related properties. Consequently, predicting impact sensitivities with reasonable accuracy appears to be too subtle a problem to be explained by a model based on fallhammer measurements alone.
Additional key factors responsible for the sensitivity of an energetic material include the molecular properties related to the kinetics and the thermodynamics of the decomposition reactions. Numerous studies have been carried out in order to correlate the impact sensitivity with properties like heat of detonation [16,17,18], detonation velocity [19], bond dissociation energy [22, 23], oxygen balance [24], electrostatic potential of the molecular surface [25,26,27], band gap [28],15N NMR chemical shift [29], “doorway modes” in the region 200–1000 cm− 1 [30], and free space in the crystal lattice [31]. More recent studies focusing on physical factors report that there are only weak correlations or trends between the impact sensitivity and heat of detonation, electrostatic potential, and free space in the crystal lattice [7, 8, 17].
Due to the complexity of initiation of the decomposition pointed out by Dlott, care must be taken before drawing mechanistic conclusions based on simple correlation studies [32]. Moreover, if a study is based on too few compounds to make conclusive judgments, we risk asserting accidental correlations [33]. Using a large set of molecules, Keshavarz et al. derived models based on the CHNO ratios and different molecular moieties [34, 35]. Quantitative structure-property relationship (QSPR) models have also been developed for large sets of molecules [36, 37], and seem to be able to predict the impact sensitivity with reasonable accuracy. However, unlike models based on physical factors, these QSPR models do not reveal much information about the intrinsic factors that govern the impact sensitivity. Since they generally contain a surfeit of adjustable parameters, they are also prone to over-fitting.
Models for predicting the impact sensitivity based on the Arrhenius equations and the thermodynamics of the decomposition were introduced in the 1940s and 1950s [12]. In a study of 15 molecules, Wu et al. showed that the ratio between the dissociation energy of the weakest –NO2 bond and the heat of decomposition correlated with the impact sensitivity [1, 44]. This approach was refined by Mathieu et al., investigating models based on larger sets of molecules [38,39,40]. For various families of energetic materials, they found correlations between the impact sensitivity and the bond dissociation energy divided by the decomposition energy. In these models, the enthalpies of formation were either neglected or calculated with a simplified method, and the decomposition energies were computed assuming that the energetic molecule decomposed to H2O–CO2 arbitrarily [38,39,40]. Instead of calculating the bond dissociation energies for each molecule separately, the –NO2 bond dissociation energy was assigned to a constant value by considering into which family of energetic materials the molecule belonged, along with the functional groups in the neighboring position of the nitro group [38,39,40].
In this work, we report density functional theory (DFT) calculations of the –NO2 bond dissociation energies and thermodynamic calculations of the heat and temperature of detonation for 70 energetic molecules. We then apply our results to investigate how these properties can predict impact sensitivity. All –NO2 bond dissociation energies in the molecules are calculated separately. When overlap between the data sets is accounted for, we reach a total of 91 data points on which our regression models are based. Impact sensitivity models based on bond dissociation energies, heats of detonation, detonation temperatures, and total energies are evaluated for these molecules. 1,3,5-Triamino-2,4,6-trinitrobenzene (TATB) is known for its low sensitivity, and has been frequently used in models for predicting impact sensitivities [18, 19, 34, 37, 38]. However, its critical impact level (490 cm;2.5 kg drop weight) is not a measured value but an estimate based on extrapolating the measured critical heights and oxygen balances of only three energetic molecules. The measured critical height of TATB is only reported to be higher than 320 cm [41]. We will therefore use our most promising model to make a more accurate prediction of this value.
Theory and methods
Modeling the critical impact level
We model the critical impact level of an energetic mole-cule as a continuous random variable I with the property that its (natural) logarithm \(\log I\) is governed by a normal distribution with mean μ and variance σ2, so that
In general, μ will depend on the individual choice of molecule, whereas σ2 is assumed to be constant across families of molecules. The variance will depend on the level of statistical noise in the data set under consideration, which is largely due to experimental inaccuracies. However, when modeling quantum mechanical phenomena, genuine randomness in nature may also have an effect on the measurements. The Bruceton method and UN test procedure are among the most common schemes for measuring critical impact levels. The former gives the impact energy level I50 (J) or height H50 (cm) at which 50% of the test samples are expected to explode, whereas the latter gives the impact energy level I1:6 (J) or height H1:6 (cm) that results in an explosion for at least one in six test drops [42, 43]. For dimension reasons, we introduce a reference value I0 of 1 J. We let the critical impact level I denote either I50/I0 or I1:6/I0, depending on whether the data set upon which the regression is based is in accordance with the Bruceton method or the UN test procedure.
We now motivate our choice of models. According to the hot-spot theory, when an energetic material is subjected to a mechanical impact, material deformation will increase the hot-spot temperatures. If this temperature is above a critical level, the molecule will decompose. The trigger linkage hypothesis states that the first step in the initiation of an energetic molecule is a bond cleavage. The decomposition is triggered by the homolytic fission of an A–NO2 bond, and so the reaction is given by
where the star ⋆ denotes the transition state and A is either C, N, or O.
At high temperatures, similar to what the material is exposed to by impact or shock, the C–NO2 bond dissociation is the dominant reaction in the initial decomposition phase for nitroaromatic molecules [33]. However, at lower temperatures, reactions involving the other functional groups on the aromatic ring may occur. Furthermore, auto-catalyzed reactions and self-heating of the material because of exothermal reactions determine the rate of reaction r in the next decomposition phase. The kinetic theory attributed to Arrhenius dictates that r is inversely exponentially dependent on the activation energy Ea (the energy required to transform the reactant into the transition state), giving us
where \(\left [\text {A--NO}_{2}\right ]\) is the molar concentration of A–NO2, t is time, c is a constant (the pre-exponential factor), n is the reaction order, R is the molar gas constant, and T is the absolute temperature. Since it has been assumed that the rate of reaction given by the Arrhenius equation correlates negatively with the impact sensitivity of an energetic material [1, 12, 38, 44], we make the assumption that these quantities are inversely proportional. That is, the sum \(\mu + \log r\) is constant. Combining this with taking logarithms on both sides of Eq. 3, we get that
When \(\left [\text {A--NO}_{2}\right ]\) and n are assumed to be constant, Eq. 4 takes the form
where c1 is a constant.
Equation 5 forms the basis for our models. In order to calculate μ, we first need to calculate Ea and T. Unfortunately, these parameters are difficult to determine. Even though Ea can be calculated by quantum mechanical methods, several transition states and different decomposition routes need to be considered, making the calculations very time-consuming. For this reason, we have evaluated various approximation schemes for Ea and T. In our first three models, we assume that the activation energy in Eq. 5 is constant.
The Arrhenius law requires a particular temperature. During the complex sequence of events leading to an explosion, the temperature in the surroundings of a decomposed molecule will deviate from the ambient temperature. A small number of neighboring molecules are envisioned to decompose and release energy, increasing the local temperature. If the decomposition reaction produces more heat than is lost in conjunction with the heating of the nearby species, heat convection, and conduction to the surroundings, the temperature will rise rapidly. If this occurs, the reaction may propagate into an explosion. Therefore, T is a local temperature, varying in space and time during these events. In this picture, the more energy released during decomposition, the higher the local temperature. Consequently, the heat of detonation Q (kJ dm− 3) is assumed to be proportional to T. This gives our first model, which was studied in References [1, 7, 8, 16,17,18, 44], namely
Here, the constants c1, c2 are fitted to critical impact level data by linear regression and Q is calculated in EXPLO5 [45], which uses the chemical formula, the enthalpy of formation, and density to calculate the detonation properties. The enthalpies of formation and densities are taken from References [46,47,48,49]. These references do not contain enthalpy of formation or density for all the energetic materials we shall consider, so for the remaining molecules, these parameters are estimated using the method described in References [50,51,52,53,54].
Song et al. assumed that the total energy Etotal (Hartree) of an energetic molecule correlates with the energy release of the decomposition reaction [20, 21]. This motivates the assumption that Etotal and T are proportional, giving us the model
where Etotal is calculated by Gaussian09 [55]. The zero-point energy is not included in this calculation.
A new model, to our knowledge not discussed in the literature, results from assuming that T in Eq. 5 is proportional to the detonation temperature Tex (K) of the material, leading us to
where Tex is calculated in EXPLO5.
The models above require a constant Ea for the homolytic cleavage of the A–NO2 bond for each class of energetic material. This is clearly a rough approximation. It has been proposed that Ea is proportional to the bond dissociation energy BDE (kJ mol− 1), that is, the energy required to break the trigger-linkage [1, 38,39,40, 44]. However, a proportional relationship between BDE and Ea is likely to hold only for compounds where the resonance stabilization and the structure of the transition states are relatively similar. Khrapkovskii et al. reported a significant correlation between the measured value of Ea for the C–NO2 homolysis in nitroaromatic substances with different substituents and the values of BDE calculated by the hybrid DFT functional B3LYP and a small basis set 6-31G(d,p), with a coefficient of determination R2 of 0.72 [56]. This motivates the assumption that Ea and BDE are proportional.
Our fourth model in question was studied in References [22, 23]. It is based on BDE alone, and reads
where we have assumed that T in Eq. 5 is constant. By including BDE and the approximations used in Eqs. 6, 7, and 8 for the local temperature into Eq. 5, we arrive at our final three models of consideration, which take the form
Equations 10 and 11 were studied in References [1, 19, 38,39,40, 44] and [20, 21], respectively.
We study three families of energetic molecules: nitroaromatics, nitramines, and nitrate esters. For each family, we make a choice of which bond rupture we believe to be the key step in the initiation process. We choose C–NO2, N–NO2, and O–NO2 for nitroaromatics, nitramines, and nitrate esters, respectively.
Density functional theory calculations
Our original intention was to optimize the geometry of molecules and radicals with the M06 functional and the 6-311+G(2d,p) basis set, since M06 is reported to calculate homolytic dissociation of C–NO2 bonds accurately [57]. However, we were not able to calculate several hundred values of BDE with this choice of functional and basis set due to limited computer power. In order to avoid this difficulty, we instead chose the B3LYP functional, which is widely used in optimizing the geometry of energetic materials. This functional is known to systematically undershoot the value of BDE for C–NO2 [57], but Khrapkovskii et al. have shown that with a small basis set (6-31G(d,p)), it calculates BDE for substituted nitroaromatics with similar accuracy as wB97xd/6-31+G(2df,p), G2, G3, G3B3 and CBS-QB3 [56]. In Table 1, we show how the calculated value of the C–NO2 BDE for nitrobenzene converges by increasing the size of the basis set. This value has been measured to be 298.7 kJ mol− 1 [58] and 314.5 kJ mol− 1 [59].
The BDE values in Table 1 are calculated according to the method in Reference [60]. Table 1 also illustrates how the B3LYP functional undershoots the BDE values, but it should be borne in mind that in our models, differences in BDE are more important than the particular values they take. The calculation is defined by
where \(E\left (\mathrm {A}\cdot \right ), E\left (\text {NO}_{2}\cdot \right )\), and \(E\left (\text {A--NO}_{2}\right )\) denote the ground state electronic energies (open shell model) of the species A⋅,NO2 ⋅ and A–NO2, respectively. In Eq. 13, the zero-point energy is neglected since Song et al. have shown that it bears no important role for the correlation between I50 and BDE/Etotal [20].
Experimental measurements
Since variation in the measurements of critical impact level causes difficulties in parametrizing and validating models, it is important to keep observational uncertainties to a minimum when conducting experiments. The critical impact level data for the nitroaromatic materials is obtained from the Wilson et al. data set, where the tests were performed in the same laboratory with the same equipment and according to the same test procedure [48]. Wilson et al. also ensured that the molecules considered had a similar particle distribution, and so we find this data set to be the most useful one for our purposes. It should be noted that the critical height of 1,3-diamino-2,4,6-trinitrobenzene (DATB) and TATB are only given as a lower limit below which they did not explode, namely H50 > 200 cm for both molecules. Therefore, these values are not included in the training set from Wilson et al. In our analysis, we also consider two other data sets for nitroaromatics in order to reduce the risk of asserting any accidental correlations. These are taken from Storm et al. [41] and Meyer et al. [47]. The critical impact levels for the nitramines and nitrate esters are obtained from Storm et al. and Meyer et al., respectively. For the Wilson et al. and Storm et al. data sets, the Bruceton procedure was used, while the Meyer et al. data set is based on the UN test procedure. Hence, there are systematic differences between the measured critical impact levels in these data sets, which emphasizes that we cannot easily combine them in order to parametrize and validate the models collectively [8].
Statistical analysis
In order to evaluate the predictive ability of our models, we calculate the coefficient of determination (R2), root-mean-square error (RMSE), absolute mean, and maximum deviation between our predictions and the measurements from our data sets. The most promising model is also developed in parallel via Bayesian regression, taking model complexity into account. We evaluate the predictive power of this Bayesian model via the model evidence function. In addition, we also perform a simple sensitivity analysis in order to evaluate the consequences of inaccurate calculations or measurements.
In the frequentist framework, the mean μ and variance σ2 in the distribution (1) are estimated using the sample mean and correctly scaled sample variance, respectively, which are unbiased. We also note that our assumption (1) may be rephrased as
and so in the frequentist framework, we may predict the mean and variance of a new critical impact level Inew as
In the Bayesian framework, we introduce a conjugate prior distribution for the model coefficients and estimate the variance by maximizing the evidence function. Then any new critical impact level \(I_{\text {\scriptsize new}}|\mathcal {D}\) given observed data \(\mathcal {D}\) will be governed by the predictive distribution, which will also be lognormal due to the functional form of the prior distribution. That is,
where \(\mu _{\mathcal {D}}\) and \(\sigma ^{2}_{\mathcal {D}}\) are the mean and variance of the predictive distribution, respectively. Recall that unlike in the frequentist framework, \(\sigma ^{2}_{\mathcal {D}}\) will depend on the individual choice of molecule considered. The expectation and variance of the new critical impact level \(I_{\text {\scriptsize new}}|\mathcal {D}\) may then be calculated as
Later, we predict the critical impact level for DATB and TATB, both in the frequentist and Bayesian framework.
Results and discussion
The molecular structure of these energetic materials together with the computed data are available in the Electronic Supplementary Material.
Nitroaromatics
We perform the linear regression for nitroaromatic materials in two separate ways, both in a frequentist framework and via Bayesian regression. In the former framework, the model coefficients are determined by minimizing a non-regularized sum-of-squares error function, and the predictive ability of the model is evaluated via cross-validation. Note that with this approach, the issue of model complexity is not addressed. However, when performing regression with relatively few data points, the issue of model complexity becomes a key point for avoiding over-fitting. In the Bayesian framework, analysis of model complexity is built in by design, which leads us to somewhat different conclusions than those relying on the frequentist approach, in particular for the Storm et al. data set. In the Bayesian framework, the predictive quality of the model is addressed by evaluating the model evidence rather than cross-validation.
Frequentist framework
The results for the nitroaromatic materials based on the Wilson et al., Storm et al., and Meyer et al. data sets are shown in Table 2.
We note that 1/Tex correlates better with \(\log I\) than 1/Q for all three data sets. The most promising predictor of critical impact levels overall is BDE/Tex, with R2 = 0.81,0.67, and 0.69. Figure 1 illustrates the merit of this model for the Wilson et al. data set.
Table 2 shows only a weak correlation between BDE/Etotal and \(\log I\). These results are not in line with those of Song et al., which indicate merit for this model [20, 21]. However, they derived this correlation by using a small data set. For molecules of similar structure, Etotal is likely to correlate with Q, but this is unlikely to hold in general; we get virtually no correlation (R2 = 0.03) when plotting these parameters against each other.
As there are only 24, 17, and 16 molecules in the Wilson et al., Storm et al., and Meyer et al. data sets, respectively, our regression is a priori prone to over-fitting. In order to evaluate its predictive ability, we use leave-one-out cross-validation; the results of which are summarized in Table 3.
Our cross-validation does not show any particularly convincing results for either data set, revealing that more parameters than just BDE/Tex are needed to predict the critical impact level of nitroaromatic compounds. The most promising numbers are for the Wilson et al. data set, with an RMSE of 6.1 J and an average absolute deviation of 3.7 J between the predicted and measured I50. 5,7-Diamino-4,6-dinitrobenzofuroxan (CL-14) has the largest deviation (25 J). The (Wilson et al.) model predicts the critical impact level I50 of CL-14 to be 48 J, while the measured I50 is 29 J. We note that the calculated C–NO2 BDE in CL-14 is quite large, at 318 kJ mol− 1. However, this NO2 group is surrounded by an amino group on each of the neighboring carbons, as is also the case for TATB. Hence, this value for CL-14 is not particularly surprising, as the BDE of TATB is calculated to be 310 kJ mol− 1. The deviation between the measured and predicted critical impact level of CL-14 may indicate that this molecule follows another decomposition route.
CL-14 contains a furoxan ring, as is also the case for three other molecules in the Wilson et al. data set, namely 7-amino-4,6-dinitrobenzofuroxan (ADNBF), 4,6-dinitrobenzofuroxan (DNBF), and 8-amino-7-nitrobenzobisfuroxan (CL-18). These have deviations of -3 J, -8 J, and 2 J (respectively) between the predicted and measured I50. If the decomposition is initiated in the furoxan ring, a larger deviation may be expected.
From Table 3, we see that in the Storm et al. data set, 1,3-dihydroxy-2,4,6-trinitrobenzene (styphnic acid) has the largest deviation between the predicted and measured I50. Our model predicts this value to be 45 J, while it is measured to be 11 J. The BDE is calculated to be 287 kJ mol− 1 by using the M06 functional and the 6-311G+(2d,p) basis set. This is similar to the value calculated when using B3LYP/6-31G(d), which is 274 kJ mol− 1.
Bayesian regression
We now perform Bayesian linear regression separately on the three data sets for the most promising model, namely (12). In order to adapt the Bayesian framework, we introduce a zero-mean, isotropic bivariate normal prior distribution over the model coefficients c = (c1, c2)T with covariance matrix τ2I, so that
We follow the process outlined in Reference [61], in which the first step is to maximize the evidence function in order to obtain estimates for the parameters σ2 and τ2, a technique also known as emperical Bayes. We then compute the posterior distributions for the model coefficients by updating our prior distributions separately over the the data sets. Using the mean of this posterior distribution as our estimate for the model coefficients c, we obtain our linear models. Note that this process is equivalent to minimizing a regularized sum-of-squares error function with regularization term λ = σ2/τ2, and so model complexity is intrinsically accounted for. A summary of the model coefficients computed, along with the relevant parameters, is provided in Table 4.
We see that the penalizing regularization term plays a substantial role for the regression on the Storm et al. data set. Although values of R2 in the frequentist analysis for Storm et al. and Meyer et al. were relatively close in value (0.67 and 0.69, respectively), the Bayesian analysis suggests that there is little evidence supporting a linear term in the model for the Storm et al. data set. Considering the interpretation of the parameter γ = cTc/τ2 provided in Reference [62], namely as the effective number of parameters for the model, we see that for the Storm et al. data set, there is effectively only a single parameter (the constant term) governing the distribution of data points. This is in contrast to the corresponding results for the Wilson et al. and Meyer et al. data sets, where the effective number of parameters is calculated to be approximately equal to 2, supporting the claim that \(\log I\) indeed depends linearly on BDE/Tex.
We evaluate the predictive ability of our models in the Bayesian framework by calculating the (\(\log \)) model evidence function. Since the Bayesian regression penalizes the model complexity for the Storm et al. data set, we also compare the evidence for our proposed model to that of a separate constant model which asserts no correlation between \(\log I\) and BDE/Tex. That is, our other model claims that I is governed by a lognormal distribution of the form
with constant mean μ0 and constant variance \({\sigma _{0}^{2}}\). We perform the corresponding Bayesian regression for this alternative model and compare the results to our original model (12) by calculating the Bayes factor. The results of this model comparison are presented in Table 5. A Bayesian factor larger than unity indicates preference towards the linear model, while a factor smaller than unity indicates preference towards the constant model.
From Table 5, we see that Bayesian regression on the Wilson et al. and Meyer et al. data sets prefers the linear model (12), whereas it prefers the constant model (21) for the Storm et al. data set. Hence, our prediction of the critical impact level of TATB will be based on the Wilson et al. data set rather than that of Storm et al. Note that the model comparison for the Wilson et al. data set is several orders of magnitude more decisive than for that of Meyer et al.
We make two predictions of the critical impact level I50 of DATB and TATB, one based on the frequentist framework and another from the Bayesian predictive distribution. The predictions and their uncertainties are calculated using Eqs. 15 and 16. The results are given in Table 6.
Storm et al. measured the critical impact level (H50, 2.5 kg drop weight) of TATB to be larger than 320 cm (I50 > 78 J). Additionally, they predicted H50 to be 490 cm (I50 = 120 J) based on the measured impact sensitivity and the oxygen balance for trinitrobenzene, 2,4,6-trinitroaniline and DATB [41]. Using the Wilson et al. data set, our frequentist and Bayesian prediction ofI50 for TATB are 97 J and 95 J, respectively. We note that the Bayesian prediction is less certain since the variance of the predictive distribution accounts for uncertainty related to the model coefficients, as well as noise in the data. This difference in certainty for the two models is illustrated in Fig. 2.
The frequentist and Bayesian predictions of I50 for DATB are 61 J and 60 J, respectively. In the Storm et al. data set, its measured value is 78 J [41], which is consistent with the difference between the Wilson et al. data set and the Storm et al. data set. Indeed, the former data set is more sensitive than the latter for materials with a critical impact level larger than 15 J. For more sensitive materials (I50 < 15 J), the differences in critical impact level are smaller [8]. In the Storm et al. data set, a sample mass of 40 mg was used, while in the Wilson et al. data set, it was 35 mg. It has been reported that nitroaromatics such as tetryl, 2,4,6-trinitrotoluene (TNT) and o-trinitrophenol become more sensitive when the thickness or the amount of sample is reduced [14, 15]. The smaller the sample, the more energy is released per unit volume of explosive. The Wilson et al. data set is based on a smaller sample mass than that used in the Storm et al. data set, which may be one of the main reasons for the difference in measured critical impact levels.
Model sensitivity analysis
We now perform a simple sensitivity analysis of our frequentist and Bayesian models, investigating how sensitive they are to slight changes in the input variables upon which the regression is based. From the “Bayesian regression” section, we expect the Bayesian model to be more sensitive to such alterations than the frequentist model, since the variance of the Bayesian predictive distribution also accounts for uncertainty in the model coefficients. This can be seen explicitly by observing how the predicted value of I50 for DATB changes if we over- and underestimate the values of BDE and Tex in turn in the Wilson et al. data set. When multiplying BDE and Tex by (1 + ε), where ε = -0.05,-\(0.04, \dots , 0.04, 0.05\), the frequentist prediction of I50 for DATB changes by no more than 10− 13% of its original value. On the other hand, the corresponding difference in the Bayesian prediction ranges from 10− 8% to 10− 6%.
Inaccuracies in the measured critical impact level may also affect how the model makes new predictions. We measure the sensitivity of our frequentist and Bayesian models by multiplying the values of I in the data set by (1 + ε), where ε is drawn (separately for each molecule) from a uniform probability distribution of range [-0.05,0.05]. We then calculate the absolute difference in the predicted I50 of DATB before and after perturbing the input, given as a percentage of the original predicted value of I50. Averaging over 500 such random simulations, we find the predicted I50 of DATB to deviate from the original prediction by a factor of 1.3% for both the frequentist and the Bayesian model. This result indicates that inexact experimental measurements may sully the accuracy of the model.
Nitramines
The results for the nitramines based on the Storm et al. data set are shown in Table 7. We see that neither 1/Q nor 1/Tex is strongly correlated with \(\log I\) (R2 = 0.34 and 0.44, respectively), and for the N–NO2 BDE, we get virtually no correlation with \(\log I\) (R2 = 0.04).
In addition to N–NO2 bond, some of the nitramines contain one or more C–NO2 bonds. For such molecules, the C–NO2 bond may be weaker than the N–NO2 bond when three NO2 groups are attached to the same carbon atom. This is due to repulsive forces between the NO2 groups and also their attraction to electrons in the C–NO2 bond. If the weakest bond in the model is taken to be either N–NO2 or C–NO2, there is still low correlation between \(\log I\) and BDE, indicating that molecular rearrangements and auto-catalyzed reactions play a key role in the initial nitramine decomposition. Contrary to the promising result for nitroaromatics, we only see a weak correlation between \(\log I\) and BDE/Tex (R2 = 0.41). This observation is illustrated in Fig. 3, where \(\log I\) is plotted against BDE/Tex.
We now comment on the outliers marked in Fig. 3. First, bis-(2,2,2-trinitroethyl)-nitramine has a positive oxygen balance which results in a relatively low value of Tex since the molecule does not contain enough carbon or hydrogen to utilize all the oxygen. This leads to a particularly high value of BDE/Tex, which illustrates how our model predicts compounds with a positive oxygen balance to be less sensitive to impact than what they actually are. Next, N,N’-dinitromethanediamine and N,N’-dinitro-1,2-ethanediamine contain the –NH–NO2 functional group. The values of the N–NO2 BDE for these molecules (220 kJ mol− 1 and 214 kJ mol− 1, respectively) are relatively high compared with those of nitramines, which usually range from 150 to 170 kJ mol− 1. We calculated BDE of these compounds with the M06 functional and the 6-311+G(2d,p) basis set in order to exclude the possibility for any erroneous geometry optimization caused by the B3LYP functional. The M06 functional with the 6-311+G(2d,p) basis set predicts BDE for N,N’-dinitromethanediamine and N,N’-dinitro-1,2-ethanediamine to be 244 kJ mol− 1 and 241 kJ mol− 1, respectively. Thus, the divergence of these two molecules from the regression line cannot be explained by the high BDE calculated by the B3LYP functional alone.
The initial decomposition of nitramines can take place through several mechanistic routes. For example, at least four initial mechanisms for the decomposition of 1,3,5-trinitroperhydro-1,3,5-triazine (RDX) have been suggested: N–NO2 homolysis, HONO elimination, the “tripple whammy” mechanism, and NONO isomerization. Using the couple cluster theory, the Ea value of the HONO elimination has been calculated to be lower than that of the N–NO2 homolytic reaction for RDX [63]. A recent study of the initial decomposition process of liquid RDX has revealed that HONO elimination is likely to be the major decomposition pathway [64].
In order to investigate whether HONO elimination is an alternative decomposition route for N,N’-dinitromethanediamine, we calculate the energy required to break the N–NOOH bond. The HONO BDE is calculated to be 420 kJ mol− 1. When the hydrogen atom is moved from the nitrogen atom to the oxygen atom, the length of the N–N bond decreased from 1.379 to 1.257 Å (M06/6-311+G(2d,p)). The bond becomes more like a double bond, making it unlikely to break without more molecular rearrangements.
Nitrate esters
The nitrate ester data set (from Meyer et al.) consists of both liquids and solids. Table 8 shows that when these are considered simultaneously, none of our models seems to give any satisfactory predictions. When solids and liquids are treated separately, no model shows any notable improvement for the liquids. Denisaev et al. have found a correlation between \(\sqrt {H_{50}}\) and ρQ, where ρ is the density, for liquid nitrate esters [16]. The critical impact levels of liquid nitrate esters are sensitive to experimental factors such as the presence of bubbles in the liquid which can significantly alter the impact sensitivity of a liquid explosive [12].
There seems to be a strong correlation between \(\log I\) and BDE/Tex for solid nitrate esters, but as can be seen in Fig. 4, this assertion is based on very few data points. In order to investigate whether this correlation is genuine or just a result of over-fitting, we perform a Bayesian model comparison with a constant model (which asserts no correlation between BDE/Tex and \(\log I\)) as in the “Bayesian regression” section. When model complexity is accounted for, the Bayes factor comes out to be 0.554, in favor of the constant model. Hence, more data points are needed for further investigation of the the merit of the linear model. Similar results hold for the other seemingly promising model, namely (8), with a Bayes factor of 0.552, again favoring the constant model.
Conclusion
By investigating 70 energetic nitroaromatics, nitramines, and nitrate esters, we have evaluated seven models for predicting critical impact level, the quantity from which impact sensitivity is determined. Input parameters were the molecules’ temperature of detonation, heat of detonation, and bond dissociation energy. Our regression was based on three separate data sets comprising 91 data points in total.
For the largest nitroaromatics data set, the bond dissociation energy divided by the temperature of detonation was the best predictor of critical impact level, with a coefficient of determination (R2) of 0.81. Leave-one-out cross-validation gave a root-mean-square error (RMSE) of 6.1 J and the absolute average deviation was 3.7 J between the predicted and the measured values. A separate Bayesian regression also assigned similar merit to the predictive power of this model, also when accounting for model complexity. The frequentist and Bayesian models predicted the critical impact level of TATB to be 97 J and 95 J, respectively. Our sensitivity analysis showed this prediction to be more robust under changes in calculations of molecular parameters than measurements of critical impact levels.
For nitramines, our analysis showed that the temperature of detonation is a moderately useful predictor of critical impact level, unlike the N–NO2 bond dissocation energy, which we found to have virtually no such predictive ability. Hence, N–NO2 homolysis is unlikely to be the only reaction taking place in the initial phase of the decomposition.
None of the models was able to predict the critical impact level of liquid nitrate esters, but for a small data set of solid nitrate esters, there seemed to be promising results for bond dissociation energy divided by the temperature of detonation. However, our Bayesian model comparison revealed that more data points are necessary to validate this correlation.
Our results regarding nitroaromatics, nitramines, and solid nitrate esters allude to the temperature of detonation being a better predictor of critical impact level than the heat of detonation. Moreover, a high ratio of bond dissociation energy to temperature of detonation indicates low impact sensitivity, whereas a small ratio suggests that the material is highly sensitive to impact. Having evaluated the predictive power of our models, we conclude that predicting impact sensitivity of energetic materials with acceptable accuracy may require inclusion of additional parameters such as hardness, crystal defects, particle size, amount of sample, heat conductivity, and heat capacity.
References
Fried LE, Manaa MR, Pagoria PF, Simpson RL (2001) Design and synthesis of energetic materials. Annu Rev Mater Res 31(1):291–321
Politzer P, Murray JS, Seminario JM, Lane P, Grice ME, Concha MC (2001) Computational characterization of energetic materials. J Mol Struct Theochem 573(1–3):1–10
Rice BM, Byrd EF (2006) Theoretical chemical characterization of energetic materials. J Mater Res 21 (10):2444–2452
Klapötke TM (2012) Chemistry of high-energy materials, 2nd edn. Walter de Gruyter GmbH & Co KG, Berlin/Boston
Brinck T, Rahm M (2014) Theoretical design of green energetic materials: predicting stability, detection, synthesis and performance. In: Brinck T, Green energetic materials (eds). Wiley, Chichester, pp 15–43
Politzer P, Murray JS (2014) Detonation performance and sensitivity: a quest for balance. In: Advances in quantum chemistry, vol 69. Elsevier, pp 1–30
Politzer P, Murray JS (2016) High performance, low sensitivity: conflicting or compatible? Propellants, Explosives and Pyrotechnics 41(3):414–425
Politzer P, Murray JS (2015) Some molecular/crystalline factors that affect the sensitivities of energetic materials: molecular surface electrostatic potentials, lattice free space and maximum heat of detonation per unit volume. J Mol Model 21(2):25
Yan QL, Zeman S (2013) Theoretical evaluation of sensitivity and thermal stability for high explosives based on quantum chemistry methods: a brief review. Int J Quantum Chem 113(8):1049– 1061
Zeman S, Jungová M (2016) Sensitivity and performance of energetic materials. Propellants, Explosives and Pyrotechnics 41(3):426–451
Chen ZX, Xiao HM (2014) Quantum chemistry derived criteria for impact sensitivity. Propellants, Explosives and Pyrotechnics 39(4):487–495
Bowden FP, Yoffe AD (1952) Initiation and growth of explosion in liquids and solids. The Syndics of the Cambridge University Press, Cambridge
Wenograd J (1961) The behaviour of explosives at very high temperatures. Trans Faraday Soc 57:1612–1620
Afanasyev GT, Bobolev VK (1968) Initiation of solid explosives by impact. Nauka Press, Moscow
Wen Y, Long X, Xiang Y, Yu S, Dai X, Li M (2015) Mass dependent of explosion probability of RDX, tetryl, and a PBX powder in drop hammer test. Propellants, Explosives and Pyrotechnics 40(3):433–438
Denisaev A, Korsunskii B, Pepekin V, Matyushin YN (2010) Impact sensitivity of liquid explosives. Combustion, Explosion, and Shock Waves 46(1):74–80
Politzer P, Murray JS (2015) Impact sensitivity and the maximum heat of detonation. J Mol Model 21 (10):262
Rice BM, Hare JJ (2002) A quantum mechanical investigation of the relation between impact sensitivity and the charge distribution in energetic molecules. J Phys Chem A 106(9):1770–1783
Mathieu D (2017) Sensitivity of energetic materials: theoretical relationships to detonation performance and molecular structure. Industr Eng Chem Res 56(29):8191–8201
Song X, Cheng X, Yang X, Li D, Linghu R (2008) Correlation between the bond dissociation energies and impact sensitivities in nitramine and polynitro benzoate molecules with polynitro alkyl groupings. J Hazard Mater 150(2):317–321
Song XS, Cheng XL, Yang XD, He B (2006) Relationship between the bond dissociation energies and impact sensitivities of some nitro-explosives. Propellants, Explosives and Pyrotechnics 31(4):306–310
Owens F (1996) Calculation of energy barriers for bond rupture in some energetic molecules. J Mol Struct Theochem 370(1):11–16
Rice BM, Sahu S, Owens FJ (2002) Density functional calculations of bond dissociation energies for NO2 scission in some nitroaromatic molecules. J Mol Struct Theochem 583(1–3):69–72
Kamlet M, Adolph H (1979) The relationship of impact sensitivity with structure of organic high explosives. ii. polynitroaromatic explosives. Propellants, Explosives and Pyrotechnics 4(2):30–34
Murray JS, Concha MC, Politzer P (2009) Links between surface electrostatic potentials of energetic molecules, impact sensitivities and C–NO2/N–NO2 bond dissociation energies. Mol Phys 107(1):89–97
Politzer P, Murray JS (1996) Relationships between dissociation energies and electrostatic potentials of C–NO2 bonds: applications to impact sensitivities. J Mol Struct 376(1–3):419–424
Murray JS, Lane P, Politzer P (1998) Effects of strongly electron-attracting components on molecular surface electrostatic potentials: application to predicting impact sensitivities of energetic molecules. Mol Phys 93 (2):187–194
Zhang H, Cheung F, Zhao F, Cheng XL (2009) Band gaps and the possible effect on impact sensitivity for some nitro aromatic explosive materials. Int J Quantum Chem 109(7):1547–1552
Zeman S (2006) New aspects of initiation reactivities of energetic materials demonstrated on nitramines. J Hazard Mater 132(2–3):155–164
Ye S, Tonokura K, Koshi M (2003) Energy transfer rates and impact sensitivities of crystalline explosives. Combust Flame 132(1–2):240–246
Pospíšil M, Vávra P, Concha MC, Murray JS, Politzer P (2010) A possible crystal volume factor in the impact sensitivities of some energetic compounds. J Mol Model 16(5):895–901
Dlott DD (2003) Fast molecular processes in energetic materials. In: Theoretical and computational chemistry, vol 13. Elsevier, pp 125–191
Brill TB, James KJ (1993) Kinetics and mechanisms of thermal decomposition of nitroaromatic explosives. Chem Rev 93(8):2667–2692
Keshavarz MH (2013) A new general correlation for predicting impact sensitivity of energetic compounds. Propellants, Explosives and Pyrotechnics 38(6):754–760
Keshavarz MH (2010) Simple relationship for predicting impact sensitivity of nitroaromatics, nitramines, and nitroaliphatics. Propellants, Explosives and Pyrotechnics 35(2):175–181
Prana V, Fayet G, Rotureau P, Adamo C (2012) Development of validated QSPR models for impact sensitivity of nitroaliphatic compounds. J Hazard Mater 235:169–177
Wang R, Jiang J, Pan Y, Cao H, Cui Y (2009) Prediction of impact sensitivity of nitro energetic compounds by neural network based on electrotopological-state indices. J Hazard Mater 166(1):155–186
Mathieu D (2013) Toward a physically based quantitative modeling of impact sensitivities. J Phys Chem A 117(10):2253–2259
Mathieu D, Alaime T (2015) Impact sensitivities of energetic materials: exploring the limitations of a model based only on structural formulas. J Molec Graph Modell 62:81–86
Mathieu D, Alaime T (2014) Predicting impact sensitivities of nitro compounds on the basis of a semi-empirical rate constant. J Phys Chem A 118(41):9720–9726
Storm C, Stine J, Kramer J (1990) Sensitivity relationships in energetic materials. In: Chemistry and physics of energetic materials. Springer, pp 605–639
Dixon WJ, Mood AM (1948) A method for obtaining and analyzing sensitivity data. J Am Stat Assoc 43 (241):109–126
(2015) Recommendations on the transport of dangerous goods: manual of tests and criteria, 6th revised edn. UN, New York. https://doi.org/10.18356/c2b83494-en. Accessed 8 August 2019
Wu CJ, Fried LE (1998) First-principles study of high explosive decomposition energetics. In: Conference: eleventh international detonation (1998) Symposium, Snowmass, CO. August 31-September 4, 1998
Sućeska M (2010) EXPLO5, version 5.04. Zagreb
(1997) Data base of thermochemical data, 4th edn. Fraunhofer Institut für Chemische Technologie
Meyer R, Köhler J, Homburg A (2007) Explosives, 6th edn. Wiley, Weinheim
Wilson WS, Bliss DE, Christian SL, Knight DJ (1990) Explosive properties of polynitroaromatics, NXC TP 7073 Naval Weapons Center. China Lake
Adolph HG, Holden JR, Cichra DA (1981) Relationships between the impact sensitivity of high energy compounds and some molecular properties which determine their performance; N, M, and ρ0, NSWC TR 80-495. Naval Surface Weapons Center, Silver spring
Keshavarz MH (2009) Predicting condensed phase heat of formation of nitroaromatic compounds. J Hazard Mater 169(1–3):890–900
Keshavarz MH (2011) Prediction of the condensed phase heat of formation of energetic compounds. J Hazard Mater 190(1–3):330–344
Rice BM, Hare JJ, Byrd EF (2007) Accurate predictions of crystal densities using quantum mechanical molecular volumes. J Phys Chem A 111(42):10,874–10,879
Lu T, Chen F (2012) Multiwfn: a multifunctional wavefunction analyzer. J Comput Chem 33(5):580–592
Lu T, Chen F (2012) Quantitative analysis of molecular surface based on improved marching tetrahedra algorithm. J Molecul Graph Modell 38:314–323
Frisch M, et al. (2009) Gaussian 09, version A.02. Gaussian Inc., Wallingford
Khrapkovskii G, Sharipov D, Shamov A, Egorov D, Chachkov D, Van BN, Tsyshevsky R (2013) Theoretical study of substituents effect on C–NO2 bond strength in mono substituted nitrobenzenes. Comput Theor Chem 1017:7–13
Qi C, Lin QH, Li YY, Pang SP, Zhang RB (2010) C–N bond dissociation energies: an assessment of contemporary DFT methodologies. J Mol Struct THEOCHEM 961(1–3):97–100
Gonzalez AC, Larson CW, McMillen DF, Golden DM (1985) Mechanism of decomposition of nitroaromatics. laser-powered homogeneous pyrolysis of substituted nitrobenzenes. J Phys Chem 89(22):4809–4814
Pruitt CJM, Goebbert DJ (2013) The C–N dissociation energies of nitrobenzene and nitrotoluene radical anions and neutrals. Chem Phys Lett 580:21–27
Foresman JB, Frisch Æ (1996) Exploring chemistry with electronic structure methods, 2nd edn. Gaussian, Inc., Pittsburgh
Bishop CM (2006) Pattern recognition and machine learning. Springer, Cambridge
MacKay DJ (1992) Bayesian interpolation. Neural Comput 4(3):415–447
Molt RW, Watson T, Bazanté AP, Bartlett RJ, Richards NG (2016) Gas phase RDX decomposition pathways using coupled cluster theory. Phys Chem Chem Phys 18(37):26,069–26,077
Khichar M, Patidar L, Thynell ST (2018) Improvement and validation of a detailed reaction mechanism for thermal decomposition of RDX in liquid phase. Combust Flame 198:455–465
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Jensen, T.L., Moxnes, J.F., Unneberg, E. et al. Models for predicting impact sensitivity of energetic materials based on the trigger linkage hypothesis and Arrhenius kinetics. J Mol Model 26, 65 (2020). https://doi.org/10.1007/s00894-019-4269-z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00894-019-4269-z