LOGISTIC MIXTURE MODEL VS ARRHENIUS FOR KINETIC STUDY OF MATERIAL DEGRADATION BY DYNAMIC THERMOGRAVIMETRIC ANALYSIS SHORT TITLE: LOGISTIC MIXTURE MODEL VS ARRHENIUS FOR KINETIC STUDY OF MATERIALS Salvador Naya1, Ricardo Cao1, Ignacio López de Ullibarri1, Ramón Artiaga2, Fernando Barbadillo2, Ana García2 1 Department of Mathematics, University of A Coruña, Spain Department of Industrial Engineering II, University of A Coruña, Spain 2 Ricardo Cao Facultade de Informática, Campus de Elviña Universidade da Coruña 15071 A Coruña. -SPAIN Tel: +34 981 167 000 Fax: +34 981 167 160 E-mail: [email protected] LOGISTIC MIXTURE MODEL VS ARRHENIUS FOR KINETIC STUDY OF MATERIAL DEGRADATION BY DYNAMIC THERMOGRAVIMETRIC ANALYSIS Keywords: logistic regression, TGA curve, thermogravimetric experiments. Abstract In this work, an alternative method to the Arrhenius equation for thermogravimetric analysis is presented. It is based in performing a logistic regression of the raw TGA data. This model assumes that more than one physical process may be involved in each mass loss step and that each physical process may extend along all the experiment. The logistic mixture obtained explains the complete TGA trace, including as many mass loss steps as the experiment has. The typical asymptotic tendency of the mass loss steps is perfectly reproduced by the model. A discussion of the model from the statistical point of view is presented as well as a comparison with other classical models. INTRODUCTION Thermogravimetric analysis (TGA) is widely used to determine kinetic parameters for polymer decomposition. Both isothermal and dynamic heating experiments can be used to evaluate kinetic parameters. Each has advantages and disadvantages [1]. In dynamic thermogravimetric analysis, the mass of the sample is continuously monitored while the sample is subjected, in a controlled atmosphere, to a thermal program, where the temperature is ramped at a constant heating rate. Ideally, a single thermogram has been said to be equivalent to a very large family of comparable isothermal volatilization curves and, as such, it constitutes a rich source of kinetic data for volatilization [2]. The classical way to study the kinetics of these processes by TGA starts from the assumption that the mass loss follows the Arrhenius equation: ⎛ E ⎞ k (T ) = A ⋅ exp⎜ − (1) ⎟ ⎝ RT ⎠ where k, the reaction rate, depends on the temperature, T, and the activation energy, E, may be considered constant in each degradation process (that appears as a clear step in the mass trace) since the degradation mechanism is supposed not to change in a narrow range of temperatures. The constant A may be calculated from A = qts , where s is the reaction order, in the case that the kinetics follow a reaction order model. OTHER MODELS Many other models start from the Arrhenius equation, modified by Sesták-Berggren [3]: dα n p = k (α ) m (1 − α ) [− ln (1 − α )] (2) dt where n, m and p are constants. Two of the most used derivative models based on this equation are Freeman and Carroll [5] and Friedman [6]. There are also some integrate models, like Ozawa [7], Flyn [8] and the one proposed by Popescu [9], that allows for calculation of n and A from TGA data obtained at several heating rates. The method proposed by Conesa considers that some organic fractions of the sample decompose independently giving an organic residue and an inorganic fraction. This model gave good correlation with the mass loss derivative data for different rubbers [10]. The method proposed by Carrasco and Costa [11] has been successfully applied to the thermal degradation of polystyron. Although the application of these models to specific cases has been checked by detailed statistical studies, all of them are based on the Arrhenius equation and can not be generally applied to material degradations following very different kinetics. Moreover, its methodology is sometimes unease. It has been said for methods based on one simple heating rate that quite different reaction models fit the data equally well (from the statistical point of view) whereas the numerical values of the corresponding Arrhenius parameters crucially differ (Vyazovkin [12]). Its physical meaning is obscure and no predictions can be done outside the range of experimental temperature. Other authors deemed the Arrhenius model inappropriate for the calculation of kinetic parameters from non-isothermal thermogravimetric curves [13]. Moreover, arising from the Kinetics Workshop, held during the 11th International Congress on Thermal Analysis and Calorimetry (ICTAC) in Philadelphia, USA, in 1996, sets of kinetic data were prepared and distributed to volunteer participants for their analysis using any, or several, methods they wished. The results obtained by each researcher were different than the ones obtained by the others, Brown et al. [14]. All of this confirms our believing that the existing models cannot be generally applied and sometimes it is not clear which one is the best suitable to each case. That is the reason to propose an alternative model that will be described in the following sections. LOGISTIC MIXTURE MODEL PROPOSED The logistic mixture model (see Naya et al. [15] and Artiaga et al. [16]) proposes to decompose the TGA trace in several logistic functions, assuming that each of the functions represents the degradation kinetics of each component of the sample. Even in the case of homogeneous materials, it is supposed that several different structures may exist, each one following its specific kinetics that may be different from the others. In this model, it is assumed that a TGA trace may be fitted by a combination of logistic functions: k Y (t ) = ∑ wi f (a i + bi t ) i =1 et . f (t ) = 1 + et (3) where i = 1,2,..., k represent different components from the mass loss process point of view, not necessarily different chemical compounds. In order to model the mass loss along time, it is assumed that the candidate functions to estimate this quantity (t , Yi (t ) ) have to satisfy that the response, Yi (t ) , should tend to 0 as t → ∞ . It implies that the parameters bi have to be negative. When t = 0 , the function Yi (t ) has to tend to the mass of the original sample. This means that, the constants wi correspond approximately, when ai is large enough, to the mass loss of the sample in each mass loss process. These processes generally appear as clear steps of the TGA trace. The function Y (t ), that represents the overall TGA trace, may be expressed as a sum of components of the form Yi (t ) = wi f (a i + bi t ) . The constants ai and bi can be interpreted in the following manner. The values bi represent the slope of the mass steps while the ratio − ai / bi denotes the step location. The values wi account for the weight of each component in the sample. Figure 1 shows an example of a logistic mixture with k = 4 , w1 = 5, a1 = 12, b1 = −4, w2 = 4, a 2 = 14, b2 = −2, w3 = 7, a 3 = 43, b3 = −5, w4 = 1, a 4 = 16, b4 = −1. PUT FIGURE 1 ABOUT HERE Once the regression function of the TGA trace is obtained, it is immediate to compute its derivatives. Thus, for example, the first derivative of the TGA trace (DTG), which is used by many kinetic models, since it represents the mass loss rate along time, may be expressed by the following equation: k dTGA(t ) = ∑ wi bi f ' (a i + bi t ) i =1 f ' (t ) = et (1 + e ) t 2 k ⎛ t − a' i ⎞ a ⎟⎟ with a ' i = − i and A reparametrization of equation (3) gives Y (t ) = ∑ wi f ⎜⎜ bi i =1 ⎝ b' i ⎠ 1 b' i = . The value b'i represents the mass loss rate and a'i corresponds to the half bi mass loss location of the i-th step. LOGISTIC MIXTURE PARAMETRIC FITTING For fitting the data to a logistic mixture model some estimation of the parameter values in (3) is needed. This task is usually performed by using statistical software. In this case, we have used the non linear regression and derivatives packages of S-plus. We considered a non linear regression model: y i = m( x i , θ ) + ε i , i = 1,2,..., n. where the response variable and the independent variable values are represented by yi and xi , respectively, θ is the parameter vector, that will be estimated by least squares and ε i are the errors, assumed to have normal distribution, with zero mean and constant variance. The residuals of the model are defined as: ei (θ ) = y i − m( x i ; θ ), i = 1,2,..., n k The parameters of the model were estimated by minimizing ∑ e (θ ) i =1 i 2 a non linear least squares methods. The fundamentals of this method were described by Gay [17] The Levenberg-Marquardt method routine for generation of the approximation sequence to the minimum point, based in the “trust region” algorithm, was used for the computation of the parameter values that minimize the residuals squared sum. This algorithm was discussed by Chambers and Hastie [18]. Details about its implementation in S-plus are given in Dennis et al. [19]. One of the problems that appear when using this fit is to choose some starting points for the different parameters to estimate. To do this, one possibility consists in trying to estimate the inflexion point by direct observation of the TGA trace. Since this method is not easy and requires previous expertise, we propose a method based in the idea of assuming that the data follow locally a logistic regression. So it is possible to fit the function logit Yi (t ) / wi to a straight line with intercept ai and slope bi , where logit (u ) = log(u / 1 − u ) . The reason for this linear fitting is explained as follows: Yi (t ) wi exp(a i + bi t ) Yi (t ) = wi ⇒ = exp(a i + bi t ) Yi (t ) 1 + exp(a i + bi t ) 1− wi ⎛ Yi (t ) ⎞ ⎟ ⎜ ⎛ Yi (t ) ⎞ wi ⎟ ⎜ ⎟⎟ = log So logit⎜⎜ ⎜ Yi (t ) ⎟ = a i + bi t ⎝ wi ⎠ ⎟ ⎜⎜ 1 − wi ⎟⎠ ⎝ PARAMETER ESTIMATION An algorithm has been implemented for automatic calculation of the model parameters. It consists of the following steps: 1. The values of those inflection points where the second derivative changes from negative to positive are estimated. The derivative of the TGA curve is obtained by approximating, with a small enough h, the limit: lim h →0 y (t + h ) − y (t ) h The second derivative is obtained, in a similar way, starting from the first derivative. To identify the points where y ′′(t ) = 0 , a narrow band [− ε ,+ε ] is considered along the y axis, preventing to identify as inflection points values where the second derivative is zero as a consequence of the derivative estimation errors. The middle points of the range where the second derivative crosses the band [− ε ,+ε ] are finally selected. 2. The values w1, 0 , w2,0 ,..., wk , 0 are computed using the difference between the mean values on the y-axis of consecutive inflection points, previously calculated in Step 1. These values visually reproduce the fall steps of the TGA curve. Denoting by m j with j = 1,2,..., k − 1 , these middle points, the w j , 0 can be obtained as: w1,0 = max{Yi / i = 1,2,..., n}− m1 w j ,0 = m j −1 − m j , for j = 2,3,..., k − 1 wk ,0 = m k −1 − min{Yi / i = 1,2,..., n} 3. Each mass loss step is detected by defining y j (t ) = y (t ) − m j for j = 1,2,..., k , in the pertaining range on the horizontal axis. 4. In order to find the values a j , 0 and b j , 0 , a straight line is fitted to the points (t i , logit ( y j (t i ) w j )) in each fragment of the TGA curve corresponding to the fall steps previously identified: ⎛ y j (t ) ⎞ ⎟ = a j + bjt logit ⎜ ⎜ w ⎟ j ⎝ ⎠ Once the initial values of the parameters have been obtained, they will be optimized using the Levenberg-Marquardt algorithm. HYPOTHESIS TEST An important issue when fitting a logistic mixture to the TGA curve is to determine the number of components in the model. To answer this questions a hypothesis test view is adopted. Starting from a k+1 components model Yi = y (t i ) + ε i k +1 y (t i ) = ∑ w j f (a j + b j t i ). j =1 we consider the null hypothesis: H 0 : wk +1 = 0 , which means that the maximum number of components of the process is k, since the weight of a (k+1)-th component would be wk +1 = 0 . The alternative hypothesis is: H 1 : wk +1 ≠ 0 , i.e., the process needs the (k+1)-th component to be explained. We fixed the level α = 0.01 and have chosen the Average Squared Error (ASE) as a test statistic. The test will be performed through these steps: 1. Initialize the number of logistic components k = 1 . 2. Use the data (t i , Yi ) to estimate the model parameters θ = ( w1 , a1 , b1 , w2 , a 2 , b2 ,..., wk , a k , bk ) by means of the Levenberg-Marquardt algorithm. 3. Calculate the ASE for the model with k components: ASE ( k ) ( k 1 n ⎛ = ∑ ⎜⎜ Yi − ∑ wˆ j f aˆ j + bˆ j t i n i =1 ⎝ j =1 ) ⎞ ⎟ ⎟ ⎠ 2 where θˆ = ( wˆ 1 , aˆ1 , bˆ1 , wˆ 2 , aˆ 2 , bˆ2 ,..., wˆ k , aˆ k , bˆk ) is the parameter values estimated in the previous step. 4. Repeat B = 500 times the following mechanism: ( ) 4.1. Draw bootstrap resamples t i , Yi* , for i = 1,2,..., n, as it will be detailed later on. 4.2. Use the bootstrap resamples to estimate the parameters replication θ * = ( w1* , a1* , b1* , w2* , a 2* , b2* ,..., wk* , a k* , bk* ) by Levenberg-Marquardt. 4.3. Calculate the bootstrap version of ASE: ASE(*k ) = 1 n ⎛ * k * w j f a *j + b *j t i ∑ ⎜ Yi − ∑ n i =1 ⎜⎝ j =1 ( )⎞⎟⎟ ⎠ 5. The p-value is approximated by bootstrap using the proportion: p − value = { ( j) 1 # j / ASE * ( k ) > ASE ( k ) B } 5.1. If p − value > 0.10 , we accept H 0 , that the number of logistic components of the process is k. 5.2. If p − value < 0.01 , the number of components, k, is increased in one unit (we reject H 0 ) and come back to Step 2. 5.3. If 0.01 < p − value < 0.10 , the result is shown and let the user to choose between accepting the model or looking for more complex one. Obtaining bootstrap resamples In order to simulate the bootstrap resamples t i , Yi* , for i = 1,2,..., n , a random error will be added to the fitting of the experimental trace obtained in Step 2: ( k ( ) ) Yi * = ∑ wˆ j f aˆ j + bˆ j t i + ε i* j =1 The random errors will be generated following these steps in order to incorporate the sample autocorrelation: 1. Denoting by Ei the difference between consecutive Yi for i = 1,2,..., n − 1 : Ei = Yi +1 − Yi The quantities E and σˆ E2 are computed: E= 1 n −1 ∑ Ei n − 1 i =1 σˆ E2 = 1 n 2 ∑ Ei − E n − 1 i =1 2. Next, the values ε i* will be obtained for i = 1,2,..., n by ε i* = ρˆ E ε i*−1 + ai where 1 n−2 ∑ Ei Ei+1 n − 2 i =1 ρˆ E = 1 n −1 2 ∑ Ei n − 1 i =1 and ai is a sequence of independent and identically distributed random variables d normal ( ai = N (0, σˆ a2 ) ) with σˆ a2 = σˆ E2 (1 + ρˆ E ) . This dependence structure assumes that the instrument measurement error follows a first order autoregressive process. See Figure 2 to observe the correlation of the measurement error. PUT FIGURE 2 ABOUT HERE Once the test was performed the required number of times, the number, k, of components needed for a correct fitting is found. Moreover, an estimation of the parameters is obtained θˆ = ( wˆ 1 , aˆ1 , bˆ1 , wˆ 2 , aˆ 2 , bˆ2 ,..., wˆ k , aˆ k , bˆk ) . PHYSICAL MEANING OF THE PARAMETERS Consider a global logistic mixture fit to a TGA curve. The amount of mass evolved in each degradation step is approximately wi . In the time axis, each process is centred at − ai bi . The value of the derivative of the i-th component at that point is 1 4 wi bi , which means that bi measures the mass loss rate, i.e. the loss speed per unit of mass with respect to time (or temperature, in experiments with a constant heating rate). CASE STUDIES In order to validate the model in extreme situations, some TGA experiments exhibiting very different behaviours were considered. The first one corresponds to the analysis of wood from Eucalyptus globulus. Wood is a very complex material, where the main components are cellulose and lignin. Its thermal behaviour is not simple and overlapping processes seem to be involved. Apparently, it decomposes into three main steps. Other complex case considered was an Epoxy-Araldite sample. Eucalyptus wood experiment In this case three logistic components were assumed. The fitting to obtain the starting values was performed in three ranges. The final values for the parameter estimates are collected in Table 1: wi i =1 13.0479 i=2 41.094 i=3 22.534 ai 5.06769 15.4589 162.17 bi -0.0113 -0.0085 -0.0856 Table 1: Parameters of the logistic mixture model for the Eucalyptus sample. Epoxy-Araldite experiment A constant was assumed in this case to represent the asymptotic value at the end of the TGA curve. Four components have been assumed. After iterating the algorithm, the fitted values of the parameters are given in Table 2, with a constant value of 0.705. wi i =1 2.404 i=2 6.971 i=3 3.075 i=4 3.448 ai 11.333 35.874 24.239 19.481 bi 0.010 0.014 0.019 0.010 Table 2: Parameters of the logistic mixture model for the Epoxy-Araldite sample. Two plots of the original TGA traces compared to the estimated functions via the logistic mixture model are given in Figures 3 and 4. PUT FIGURE 3 AND 4 ABOUT HERE COMPARISON BETWEEN THE LOGISTIC MIXTURE MODEL AND THE ARRHENIUS MODEL Since the Arrhenius model is usually applied to apparently single processes, in order to compare it with the logistic mixture model, separate single mass loss steps of a polyether-polyurethane TGA trace (Figure 5) were used for Arrhenius, while the logistic mixture was applied to the overall TGA curve. The TGA test was performed at a constant heating rate of 10 ºC/min, using 50 ml/min of Argon as purge. Three Arrhenius based methods were applied: linear regression with the standard Arrhenius equation, Freeman and Carroll and Sestak-Berggren. Table 3 shows the fitted parameters as well as the correlation coefficient, r, obtained with the three Arrhenius methods in the range from 2250 to 2808 s. Since the best correlation coefficient resulted from the linear regression Arrhenius, this method was applied to all the apparently single steps of the curve. Table 3 shows the p-values and ASE obtained with the logistic mixture for 1, 2, 3, 4 and 5 components. The later resulting to be the optimal fitting. Table 4 shows the ASE obtained in the Arrhenius and the logistic mixture cases in the time ranges where Arrhenius model was fitted and Table 5 shows the p-value and the ASE obtained in the hypothesis test of the polyether-polyurethane case. As it can be seen in Table 4 the performance of the logistic mixture fit is good independently of the range considered, which is not the case of the Arrhenius method. Arrhenius-linear regression Freeman and Carroll Sestak-Berggren -16 +/- 821 kJ/mol 494 +/- 115 kJ/mol Ea -497 +/- 38.2 kJ/mol 1.475 +/- 0.025 4.310 +/- 004 -43.516 +/- 11.159 n -194.570 +/- 46.678 M 393.53 +/- 90.573 P 1.4E-99 A 2.8E-34 0.9781 0.0011 0.977 r Table 3. Parameter values and correlation coefficient obtained from a polyetherpolyurethane TGA trace for three Arrhenius based models. Time range/s Arrhenius Logistic mixture 1405.5-1685.5 0.00058262 0.00039103 1681.5-1921.5 0.246024348 0.000442474 1405.5-1861.5 0.00806295 0.00043081 1801.5-2133.5 0.00026211 0.00020021 Table 4. ASE values obtained, in the specified ranges, with the Arrhenius and logistic mixture models. ASE Number of p-value components 1 0 0.202061 2 0.01 0.011231 3 0.004 0.004747 4 0.004 0.004570 5 1 0.000256 Table 5. p-values and ASE obtained in the hypothesis test for the logistic mixture model in the polyether-polyurethane case. wi i =1 3.158 i=2 0.686 i=3 1.296 i=4 1.714 i=5 0.490 ai 51.812 207.606 47.998 24.057 8.208 -0.026 -0.010 -0.003 bi -0.031 -0.124 Table 6: Parameters of the logistic mixture model for the polyether-polyurethane sample. PUT FIGURE 5 ABOUT HERE CONCLUSIONS 1. The logistic mixture model allows for including at once the overall trace from a TGA experiment, while the classical methods can only be applied to a single step each time. 2. Overlapping degradation processes can be explained by the new method. Since the existing models were proposed to explain single processes, they generally fit very poorly to overlapped processes. 3. The thermal degradation of each component of the sample can be explained, through the logistic mixture model, by a single function that may be easily understood from the physical point of view. 4. This model shows the contribution of each single degradation process to the overall curve. It is very useful in order to improve thermal stability of materials. 5. It allows for measuring the statistical goodness of fit by signification tests. 6. The classical kinetic models are easier to apply on the estimated functions obtained by the new method than on the raw TGA data, since the raw data content noise that affect derivative estimation. 7. The asymptotics are perfectly reproduced at the beginning and end of each degradative process. ACKNOWLEDGEMENTS The authors gratefully acknowledge for the MEC Grant MTM2005-00429 (European FEDER support included), for the first four authors and XUGA Grant PGIDT03PXIC10505PN, for the first three authors. REFERENCES 1. Sircar A. K. Elastomers in Thermal Characterization of polymeric materials Vol 1, 2nd Ed. Turi E. A., Academic Press, San Diego. USA. 1997; p. 1254. 2. Doyle C. D. Kinetic analysis of thermogravimetric data. Journal of Applied Polymer Science 1961; 15, 285-292. 3. Sesták J., Berggren G. Study of the kinetics of the mechanism of solid-state reactions at increasing temperatures. Thermochimica Acta 1971; 3: 1-12. 4. Suwardie J. H., Artiaga R. The determination of heat of curing accompanied by reactant volatilization using simultaneous thermal analysis (STA). Thermochimica Acta 2000; 357, 205-210. 5. Freeman B. and Carroll B. The application of thermoanalytical techniques to reaction kinetics. The thermogravimetric evaluation of the kinetic of the decomposition of calcium oxalate monohydrate. Journal of Physical Chemistry 1958; 62: 394-397. 6. Friedman H. L. Kinetics of thermal degradation of char forming plastics from thermogravimetry. Application to a phenolic plastic. Journal of Polymer Science 1964; Part C, 6: 183-195. 7. Ozawa T. A New Method of Analyzing Thermogravimetric Data. Bulletin of the Chemical Society of Japan 1965; 38: 1881-1886 8. Flynn J. H.,Wall L. A., Quick A. Direct Method for the Determination of Activation Energy from Thermogravimetric Data. Polymer Letters 1966; 4: 323328 9. Popescu, C. Variation of the maximum rate of conversion and temperature with heating rate in non-isothermal kinetics. Thermochimica Acta 1984; 82: 387-389. 10. Conesa J. A., Marcilla A. Kinetic study of the thermogravimetric behaviour of different rubbers. Journal of Analytical and Applied Pyrolysis 1996; 37: 95-110. 11. Carrasco F. and Costa J. Modelo Cinético de la descomposición térmica del poliestireno. Ingeniería Química 1989; 121-129. 12. Vyazovkin S. A unified approach to kinetic processing of no isothermal data. International Journal of Chemical Kinetics 1996; 28: 95-101. 13. Arnold M., Veress G. E., Paulik J., Paulik F. A critical reflection upon the application of the Arrhenius model to non-isothermal thermogravimetric curves. Thermochimica Acta 1982; 52: 67-81. 14. Brown M.E., Maciejewski M.,Vyazovkin S., Nomen R., Sempere J., Burnham A., Opfermann J., Strey R., Anderson H.L., Kemmler A., Keuleers R., Janssens J., Desseyn H.O., Chao L., Tong B., Roduit B., Malek J. and Mitsuhashi T. Computational aspects of kinetic analysis. Thermochimica Acta 2000; 355: 125143. 15. Naya S., Cao R. and Artiaga R. Local polynomial estimation of TGA derivatives using logistic regression for pilot bandwidth selection. Thermochimica Acta, 2003, 406, 177-183. 16. Artiaga R., Cao R., Naya S., García A. and Varela A. (). Logistic approach to polymer degradation in dynamic TGA. Polymer Stability and Degradation, 2004, 85, 667-674. 17. Gay D. M. A trust region approach to linearly constrained optimization in Numerical Analysis. Springer, Berlin, 1984, 171-189. 18. Chambers J. M. and Hastie T. J. Statistical Models in S. Pacific Grove, CA Wadsworth & Brooks, 1992, Chapter 10. 19. Dennis J. E., Gay D. M. and Welsch R. E. An Adaptive Nonlinear Least-Squares Algorithm ACM Transactions on Mathematical Software, Springer, Berlin 1981, 348-368. 20. Suwardie J. H. Artiaga R. and Barbadillo F. Simultaneous thermal analysis of hexahydrophtalic anhydride, Thermochimica Acta 2002, 392-393, 289-294. FIGURES: 18 16 14 Mass /mg 12 LC1 LC2 LC3 LC4 Mixture 10 8 6 4 2 0 0 5 10 15 20 25 Time /min Figure 1: A logistic mixture model with four components. 7.32 7.315 Mass /mg 7.31 Data Fitting 7.305 7.3 7.295 7.29 1500 1505 1510 1515 1520 Time /s Figure 2: Autocorrelation data with fitted logistic mixture model. 1525 1530 120 100 Mass /% 80 Measured LC1 LC2 LC3 fitting 60 40 20 0 0 500 1000 1500 2000 Time /s Figure 3: Plot of the original TGA trace compared to the fitted logistic mixture model for the Eucalyptus sample. 18 16 14 Measured LC1 LC2 LC3 LC4 fitting Mass /mg 12 10 8 6 4 2 0 0 500 1000 1500 2000 2500 3000 Time /s Figure 4: Plot of the original TGA trace compared to the fitted logistic mixture model for an epoxyAraldite sample. 8 7 Mass /mg 6 Measured LC1 LC2 LC3 LC4 fitting 5 4 3 2 1 0 0 500 1000 1500 2000 2500 3000 Time /s Figure 5: Plot of the original TGA trace compared to the fitted logistic mixture model for a polyether-polyurethane sample.

© Copyright 2020