Microsoft Word - 1.docx
CHEMICAL ENGINEERING TRANSACTIONS
VOL. 77, 2019
A publication of
The Italian Association
of Chemical Engineering
Online at www.cetjournal.it
Guest Editors: Genserik Reniers, Bruno Fabiano
Copyright © 2019, AIDIC Servizi S.r.l.
ISBN 978-88-95608-74-7; ISSN 2283-9216
Predictive Methods for Determining the Thermal
Decomposition Properties of Hazardous Substances
Guillaume Fayeta,*, Klaus-Dieter Wehrstedtb, Annett Knorrb, Patricia Rotureaua
a
INERIS, Parc Technologique Alata, BP2, 60550 Verneuil-en-Halatte, France
b
BAM, Bundesanstalt für Materialforschung und -prüfung, Unter den Eichen 87, 12205 Berlin, Germany
guillaume.fayet@ineris.fr
Due to the fast development and availability of computers, predictive approaches are increasingly used in the
evaluation process of hazardous substances complementary to experiments. Their use was recommended as
alternative to experimental testing by the REACH regulation to complete the lack of knowledge on properties
for existing substances that must be registered before 2018 (upon quantities). Among the proposed predictive
approaches, Quantitative Structure Property Relationships (QSPR) are powerful methods to predict
macroscopic properties from the only molecular structure of substances.
In that context, the HAZPRED project (2015-2018, founded by the SAF€RA consortium) aims to develop
theoretical models (e.g. QSPR) and small-scale tests to predict complex physico-chemical properties (e.g.
thermal stability, explosivity) of hazardous substances to complete the lack of knowledge on these hazardous
substances quickly or to understand their decomposition behaviour better.
In particular, this contribution will present the work done in this project on the physical hazards of organic
peroxides and self-reactive substances: gathering of existing experimental data, new experimental campaigns,
review of existing models and proposition of new estimation methods.
1. Introduction
Up to now, experimental characterization is used by laboratories to gather data on physico-chemical
hazardous properties of energetic substances. Due to the fact that these tests are complex, costly and require
a large amount of substances (more than 1 kg for hazardous properties such as explosivity) and dedicated
facilities (both small and large scale), alternative or predictive methods are more and more encouraged. In
particular, the development and use of QSPR models are recommended by the REACH regulation. Moreover,
QSPR or small-scale tests (requiring lower quantities of materials) present a real interest for screening
purposes in R&D development phases (even when chemicals are not yet synthesized) or when they are very
expensive or dangerous for the users (e.g. toxic). These estimation methods can be used to replace huge
experimental characterization and avoid important time and costs on potential late-stage failures.
In this context, the HAZPRED project addresses the development and use of predictive methods, in particular
QSPR and small-scale tests, for investigating hazardous properties of chemical substances in a regulatory or
risk assessment context. Two families of hazardous substances were particularly targeted, organic peroxides
and self-reactive substances, that are both thermally instable substances.
For each system, after a collection of existing data on physical hazards (from literature and new experiments)
and a review of existing prediction methods, works were initiated towards predictive approaches for the
estimation of these properties.
2. Organic Peroxides and Self-Reactive Substances
Organic Peroxides (OP) and Self-Reactive Substances (SRS) are two classes of thermally unstable
substances that can undergo self-accelerating decomposition, burn rapidly and even, in some cases, be
explosive under the effect of various types of stimuli (e.g. impact, friction). They can also react dangerously
with other substances. For this reason, these compounds are classified under specific classes in chemical
DOI: 10.3303/CET1977057
Paper Received: 2 November 2018; Revised: 24 May 2019; Accepted: 3 July 2019
Please cite this article as: Fayet G., Wehrstedt K.-D., Knorr A., Rotureau P., 2019, Implementation of Dynamic Simulation to Study the
Deviation Consequences Along a Complex Process, Chemical Engineering Transactions, 77, 337-342 DOI:10.3303/CET1977057
337
safety regulations like the UN Recommendations for the Transport of Dangerous Goods (UN, 2017) or the
European Regulation related to the Classification, Labelling and Packing of substances (CLP) (EC, 2006).
Organic peroxides are defined by their molecular structure, as organic substances (liquid or solid) that contain
O-O bonds in their structure (as shown in Figure 1). They are classified in class 5.2 in the TDG
recommendations.
Self-reactive substances are defined as thermally unstable substances liable to undergo a strongly exothermic
decomposition without participation of oxygen (air) and classified in class 4.1. In this class of compounds are
found different types of nitrogenated compounds as presented in Figure 1.
Figure 1: Different families of organic peroxides and self-reactive substances (list is not exhaustive)
3. Collection and evaluation of existing experimental data and predictive models on organic
peroxides
3.1 Collection of existing experimental data on organic peroxides
Resources of reliable data are rare in the open literature, where comprehensibility of results is partly a weak
point. One credible source, open to members of the International Group of Experts on the Explosion Risks of
Unstable Substances (IGUS), is a so called “Datatop”, collected by TNO (2013), The Netherlands. It is a
compilation of organic peroxide classification data and is published multi-annual. Another source is the data
collection of BAM owns measurements. From both sources, results of tests and classification of new pure
organic peroxides and mixtures of organic peroxides are incorporated in the Recommendations on the
Transport of Dangerous Goods (Model Regulations) (UN 2017).
Datatop and BAM collections combine nearly 400 entries for organic peroxides, whereby e.g. for one
substance more than one entry can exist. One reason is e.g. the different final classifications of diluted
peroxides depending on varying concentration, and/or on varying diluent. The properties used for classification
purposes are listed in these databases. Several tests are mandatory to classify a sample according to the
Recommendations on the Transport of Dangerous Goods (Manual of Tests and Criteria) (UN 2015). Included
are laboratory tests and tests of the sample as packaged for transport concerning specific properties like a
possible propagation of a detonation or of a deflagration. To address another important parameter for
transport, the self-accelerating decomposition temperature (SADT) must be determined. Seemingly a huge
data collection is available for developing predictive models. But it must be remembered, that the
characterization of a property isn’t done with a singular test. Different tests for the same substance property
are combined in a test series. So, the characterization of a property is not necessarily connected to a single
measured value and therefore difficult to describe in a model.
Also, the effects of varying concentrations, different diluents or additives (mostly stabilizers) on estimated
properties are manifold and for a systematic approach not sufficiently to describe. For the development and
test of predictive models at best homogeneous data are needed. Excluding e.g. dilution effects the data are
reduced to entries for technical pure substances, namely with concentrations higher than 90 % w/w.
One property which is estimated for all samples and in advance of all other tests, is the screening of the
thermal stability with Differential Scanning Calorimetry (DSC). It is a common standard test in safety labs and
provide amongst others the released decomposition energy, which is a valuable characteristic of the sample.
A combination of this characteristic, of the peroxide structure and other measurable test data seems to be a
promising projection for a descriptive model.
338
3.2 Simple correlation between DSC data and SADT
Thermal analysis methods like DSC offer the possibility to determine kinetic parameters of a reaction,
including activation energy, pre-exponential factor and reaction model. With the knowledge of the kinetics, of
the parameters of the packaging like size, heat transfer properties and physical properties of packaged
substance, which influence the heat transfer, the SADT can be calculated.
For a first assessment of SADT, at least for a hint, if SADT will be lower than 75 °C (a limit for specific
transport requirements), a faster approach would be desirable. A simple diagram, like the one published by
Yoshida (1987) for the correlation of SADT and DSC-Onset temperature for thermal instable substances,
would be of interest. Desirable would be a linear dependence for organic peroxides too and relatively good
correlation was found for 23 pure organic peroxides, as evidenced in Figure 1a.
Malow and Wehrstedt (2005) developed an approach to predict SADT for pure liquid organic peroxides, from
dynamic DSC runs. The transfer equation, based on the standard theory of explosion by Semenov and
already used for the evaluation of DTA runs (Steinbach and Klais, 1997) was adopted for DSC. Calculated
and measured SADT (UN H.4 test, UN 2015) values were in good agreement for the 10 studied organic
peroxides. In the context of the HAZPRED project this dataset has been successfully extended by
measurements of additional pure 13 peroxides (see Figure 1b), 6 of them not in the liquid but solid state at
ambient conditions. Only Di(2-ethylhexyl)peroxydicarbonate presents a significant deviation (calc. 23 °C vs.
exp. 10°C). Origin of this deviation is under study. It also may be noticed that this approach might be inefficient
for diluted peroxides.
Figure 1: Measured SADT vs. experimental DSC Onset-Temperatures (a) and calculated SADT (b).
3.3 Existing QSPR models for organic peroxides
Few QSPR models dedicated to organic peroxides were found in literature in the last decade. Lu et al. (2011)
proposed preliminary models for the decomposition energy and onset temperature of OP. Developed from a
database of only 16 organic peroxides, the authors were not able to perform any external validation. So, if
observed correlations revealed promising with R² of 0.92 and 0.96 for the decomposition energy and onset
temperature of decomposition, respectively, their real predictive powers are unknown.
More recently, within the French PREDIMOL project, INERIS used a larger set of 38 data issued from a
dedicated experimental campaign, allowing an external validation of new predictive models. New QSPR
models (Prana, 2014) were developed for the decomposition energy and onset temperature by using, notably,
descriptors related to the properties of the peroxy bond and applying a complete validation process, according
to the OECD validation principle for regulatory purpose (OECD, 2007). The performances of the models in
external validation were good with R²ext of 0.81 and 0.83 pour the decomposition energy and the onset
temperature, respectively. Nevertheless, as shown in Eq. 1 (for the heat of decomposition), the application of
these models requires quantum chemical calculations.
193712631129399054 1 −++−=Δ OOOOOO QdnCH κ (1)
339
with C the concentration in organic peroxide, 1κ the Kier shape index (order 1), nOO the number of peroxy
bonds, dOO the distance of the peroxy bonds and QOO the Mulliken mean atomic charge on the O atoms of the
peroxy bonds.
To access easier prediction, new models (Prana, 2017) were developed for the heat of decomposition
focusing on the constitutional descriptors. If their performances were slightly lower than the quantum chemical
one, they allow first estimation of the heat of decomposition based on the only knowledge of the 2D structure
of OPs. These data were also used by Zohari (2016) to develop prediction models also based on simple
constitutional descriptors with good predictive performances in prediction (R²ext=0.92), even if their validation
may be enforced considering the relatively large numbers of structural features encountered into the
descriptors included into these models compared to the numbers of compounds used to develop them.
A series of models (Pan, 2014; Gao, 2015; Wang 2016) were also proposed to predict the self-accelerating
decomposition temperature (SADT) of organic peroxides using the QSPR approach without any experimental
measurement. The best predictive performances were proposed by Wang et al. (2016) with a prediction error
of 6.4°C calculated on test set of 10 compounds. If the obtained performances are already promising, further
experimental data with more details on the tested samples and experimental protocols may be required.
Indeed, some details on the tested samples and protocols are critical for the reliability and the homogeneity of
the data in the perspective of development and validation of predictive models. Indeed, some commercial OP
are mixtures, some are strongly diluted in solvent and different protocols can be used to determine the SADT.
4. Production of new dataset and first predictive models for self-reactive substances
4.1 Collection of existing experimental data on self-reactive substances
If data availability for organic peroxides could be regarded as good (but difficult to use in detail), the data
availability for self-reactive substances and its use revealed quite more challenging.
According to the test methods, they behave very similarly to organic peroxides and are described in the same
chapter in the UN Recommendations (UN 2015). No data collection like to the one built for organic peroxides
exist. Nevertheless, some reliable data can be accessed from the UN database, Model Regulations (UN
2017), and from BAM own ones.
As already mentioned for organic peroxides, also for an individual self-reactive substance several entries can
exist due to varying concentration or addition of supplementary components. An arrangement in groups is
more difficult, even if typical structural patterns like azo- imidazole- or other groups are identified. Often more
than one characteristic group can be found in the structure. Several compounds are salts, whereby the cation
is identical, but the anion is different.
With restriction to technical pure substances about 100 data were collected, a mapping to a specific functional
group gave about 60 entries for diazonium salts. Unfortunately, there are only fragmentary data for an
individual entry. Often, measured values are given as “higher than” or “lower than”, which is difficult to
implement in a model. Additionally, basic data from DSC are frequently missing. So, these data revealed not
sufficient to access predictive methods.
4.2 New experimental campaigns on potential self-reactive substances
In this situation, the measurement of 50 technical pure, potential self-reactive substances, supplied by Bayer
AG, gave at least a reliable dataset for DSC and BAM Fallhammer Test (UN test 3a (ii)/A14; UN 2015) results.
Most compounds (29) are characterized by one functional nitro group, whereby 6 of them also contain a
pyrazole group. Five other substances have two nitro-groups each. Other functional groups are e.g. nitroso,
trazine and azo, partly combined in a structure with another functional group. The basic structure of all, except
of one compound is a five- or six-membered ring, mostly aromatic, several times with a hetero atom like N, S
or O. Rings are substituted differently. A systematic assignment to defined substance groups is not possible,
due to the presence of multiple functional groups in their structures. All compounds are solids, but not for all a
melting peak in the thermogram could be detected.
The decomposition energy of all samples covered a range from about -3400 J.g
-1 to -312 J.g-1, whereby a
positive test result in the Fallhammer Test did not depend on the height of the decomposition energy. In the
Fallhammer Test, the individual sample was exposed firstly to an impact energy of 25 J. As shown in Table 1,
only one substance showed a positive result. In a second run, the samples were stressed with 40 J. Finally,
three samples more showed a positive result. A simple correlation between Onset Temperature and
Fallhammer Test result could not be found too.
340
Table 1: Substances with positive results “+” at 25 J and 40 J, respectively according to BAM Fallhammer Test
Name 25 J 40 J
1,3-Bis-(3,4-dichlorophenyl)-2-λ-4-diazathia-1,2-diene +
4-Chloro-3-(dimethylnitroryl)benzoic acid +
2,6-Dimethyl-3-nitro-1-(2,4,6-trichloro-3-nitrophenyl)-pyridin-4(1H)-one + +
6-Hydroxy-5-nitropyrimidin-4(3H)-one +
4.3 Structure-Property analysis
No QSPR model dedicated to self-reactive substances was found in literature, probably due to the small
number of complete and fully documented experimental data, limited to few compounds (like 2,2'-
azobisisobutyronitrile (AIBN) or 2,2'-azobis-(2-methylbutyronitrile (AMBN)) in available literature. Taking
advantage of the new experiments performed within the HAZPRED project, a structure-property analysis was
initiated on thermal stability properties (onset and maximum temperatures, heat of decomposition). It may be
noticed that, for sake of homogeneity on molecular structure description, salts and hydrated structures were
excluded. Thus, the investigated database contained 45 potential self-reactive substances.
The structure of each compound was encoded into a SMILE code. These structures were loaded into Dragon
software (Kode, 2016) and 3792 molecular descriptors were calculated. Three properties were analysed: the
heat of decomposition, the onset temperature, the maximum decomposition temperature.
A partial least square (PLS) analysis was performed using SIMCA P+ software (Umetrics, 2009). This
projection method allowed a mapping of the samples in the chemical space of the database (represented by
the calculated molecular descriptors) considering the correlation of the descriptors with target properties. After
data curation (to eliminate descriptors with no (or almost no) variance in the dataset), a first PLS analysis was
performed on the 45 samples characterized by 1909 molecular descriptors and considering the three
investigated properties. Among these three properties, the heat of decomposition was found to present the
best correlation with the molecular structures of the investigated compounds with R²Y = 0.65 compared to 0.22
and 0.24 for the onset and maximum decomposition temperatures, respectively.
To improve the structure-property analysis for this property, a new PLS analysis was fitted by focusing on the
heat of decomposition. The score scattering plot of the obtained PLS (in Figure 2) illustrates the dispersion of
the molecules in the chemical space and the correlation of this chemical space with the properties (from blue
for the lowest ΔH to red for the highest ones). The final PLS model estimates the heat of decomposition with a
correlation R² of 0.80.
Considering these preliminary results, further works are in progress to develop new QSPR models for this
property using a full validation process including an external validation to estimate their predictive capabilities.
Figure 2: Representation of the potential self-reactive substances in the PLS model obtained for the heat of
decomposition
341
5. Conclusions
The in-depth review of available data on organic peroxides and self-reactive substances illustrated well the
importance of data curation when looking at the hazardous properties of reactive substances and, in
particular, in the perspective of prediction methods.
If first QSPR models were already proposed in literature for organic peroxides, the HAZPRED project
encouraged the follow up of these research by improving data curation to improve the development and
validation of models and by extending the availability of robust and homogeneous data.
In this perspective, new experimental data were generated on the thermal stability and impact sensitivity of a
series of potential self-reactive substances. The ongoing work towards structure-property trends on the
thermal stability already provided promising results. Validated QSPR models are now expected, in particular
for the heat of decomposition from the new data obtained from the new experimental campaigns conducted in
this project using homogeneous protocols and with complete documentation on the tested samples.
Acknowledgments
The HAZPRED project is a collaboration between INERIS (France), BAM (Germany) and Technical University
of Ostrava (Czech Republic) under the SAF€RA ERA-NET collaboration (call.safera.eu). INERIS thank the
French Ministry for an Ecological and Solidary Transition for its financial support on this project. Authors thank
Bayer AG for the supply of the samples tested in the new experimental campaigns.
References
Kode, 2016, Dragon (software for molecular descriptor calculation) version 7.0.6, Kode srl, https://chm.kode-
solutions.net.
EC, 2006, Regulation N°1907/2006 of the European Parliament and of the Council of 18 December 2006
concerning the Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH).
Lu Y., Ng D., Mannan M.S., 2011, Prediction of the Reactivity Hazards for Organic Peroxides Using the QSPR
Approach, Industrial Engineering Chemistry Research, 50, 1515-1522.
OECD, 2007, Guidance Document on the validation of (Quantitative) Structure-Activity Relationships [(Q)SAR]
models, Paris.
Malow M., Wehrstedt K.-D., 2005, Prediction of the self-accelerating decomposition temperature (SADT) for
liquid organic peroxides from differential scanning calorimetry (DSC), J. Haz. Mat., A120, 21-24.
Prana V., Rotureau P., Fayet G., Andre D., Hub S., Vicot P., Rao L., Adamo C., 2014, Prediction of the
thermal decomposition of organic peroxides by validated QSPR models, Journal of Hazardous Materials,
276, 216-224.
Prana V., Rotureau P., Andre D., Fayet G., Adamo C., 2017, Development of Simple QSPR Models for the
Prediction of the Heat of Decomposition of Organic Peroxides, Molecular Informatics, 36, 1700024.
Steinbach J., Klais O., 1997, Prediction of Self-Reactive properties from DTA-Measurements, HAZARD XIII
Symposium, Process Safety-The Future, Manchester, ICHEME Symposium Series No. 141, 261-272.
TNO, 2013, Datatop99, Datatop2005 and Datatop2013, TNO Defence, Security and Safety, Rijswijk, The
Netherlands.
Umetrics, 2009, SIMCA P+, Version 12.0, Umetrics AB.
UN, 2015 Recommendations on the Transport of Dangerous Goods, Manual of Test and Criteria, 6
th rev. ed.,
United Nations, New York and Geneva, 2015.
UN, 2017, Recommendations on the Transport of Dangerous Goods, Model Regulations, Vol. 1, 20th rev. ed.,
United Nations, New York and Geneva.
Zohari N., Keshavarz M.H., Dalaei Z., 2016, Prediction of decomposition onset temperature and heat of
decomposition of organic peroxides using simple approaches, Journal of Thermal Analysis and
Calorimetry, 125, 887-896.
Yoshida T., 1987, Safety of Reactive Chemicals. Industrial Safety Series 1., Elsevier, Amsterdam-Oxford-New
York-Tokyo.
342