Spatial-temporal semi-empirical dynamic modeling of thermal gradient CVI processes D. K. Rollins, Sr.a, b,, D. J. Rollinsc, and A. D. Jones, Jr.d a Department of Chemical and Biological Engineering, Iowa State University, Ames, Iowa, USA b Department of Statistics, Iowa State University, Ames, Iowa, USA c Materials Science and Engineering Department, The Ohio State University, Columbus, Ohio, USA d Department of Mathematics, Florida A&M University, Tallahassee, Florida, USA Abstract Thermal gradient chemical vapor infiltration (CVI) appears to have much promise as a process to produce carbon/carbon composites and has several advantages over conventional isothermal CVI. Using a graphite heater in the center of porous disk preforms, Zhao et al. (2006) reported excellent densification and no surface preform plugging. The aforementioned work produced a large amount of spatial and dynamic temperature data but falls short of providing a model of temperature over space (r) and time (t). Hence, using the data reported by Zhao et al. (2006), this article develops and presents two dynamic models for deposition temperature as a continuous function of t and distance from the heater surface, r. One model uses only two adjustable parameters and the other five and both capture the phenomenological structural behavior quite well. Hence, the approach presented in this work provides a methodology to model real processes using a limited amount of data and produces important continuous dynamic and spatial behavior that can prove valuable in controlling and optimizing thermal gradient CVI processes. Keywords: Carbon/carbon composites; Chemical vapor infiltration; modeling Correspondance to: Professor D. K. Rollins, Sr., Department of Chemical Engineering, Iowa State University, Ames, IA 50011, USA. E-mail: drollins@iastate.edu 1 INTRODUCTION Chemical vapor infiltration is a relatively new process for forming carbon reinforced carbon composites. This is an important process because carbon/carbon composites have been proven to have excellent high-temperature mechanical properties and high thermal stabilities (Li et al., 2005; Zhang and Hüttinger, 2003; Vignoles et al., 2004; Delhaes, 2002; Birakayala and Edward, 2002). This material exhibits high strength, stiffness, toughness, thermal shock, wearresistance, ablation performance, friction resistance and is also lightweight (Zhang and Hüttinger, 2003). These composites are already being used where high temperatures and friction are an issue such as in brake pads for planes and on space crafts (Birakayala and Edward, 2002); unfortunately though because of high processing times this material is extremely expensive, limiting its possible usage. The conventional CVI process consists of infiltration of fiber preforms with a carrier gas such as helium and a low molecular weight hydrocarbon such as methane or ethane at constant temperature and pressure. The temperature is usually around 1000 C and the hydrocarbons react rapidly to form larger molecules that deposit into the carbon matrix at a high rate typically plugging the entrance of pores and inhibiting infiltration and thus, densification. A promising alternative to overcome the limitations of the isothermal CVI process is thermal gradient CVI (Tang et al., 2003; Golecki et al., 1995; Probst et al., 1999; Jiang et al., 2002). As the name suggest, thermal gradient CVI seeks to control carbon deposition by controlling the temperature profile in the preform. This is done by creating a higher temperature where deposition is to occur and maintaining a lower temperature where deposition is not to occur. Filling of the pores is controlled by temperature in the zones of the preform. Thermal gradient CVI was demonstrated in Zhao et al. (2006) using an electrical furnace consisting of a heated rod between two 2 electrodes with cylindrical donut shaped preforms slid over the rod. Thus, the preforms were heated in the radial direction from the center out. Natural gas was introduced into the reactor from the cold outer surface at a controlled rate and reached the hot deposition zone by diffusion. The deposition zone is controlled by temperature and a thermocouple inserted into the preforms that moves in the radial direction outward as the temperature rises and densification is completed. This work develops two types of dynamic semi-empirical models of the temperature field of the preform continuously over time and space for the process in Zhao et al. (2006) using only the data they reported. Moreover, this article will demonstrate a new model building process for this application and provide theorists and practitioners with a means to obtain accurate spatial and dynamic response surface behavior. The models give the temperature (T) as a function of time (t) and radial distance from the heater surface (r). The difference between the models is the dynamic form they use. Model 1 uses a third order, critically damped transfer function and depends only on two adjustable coefficients. Model 2 uses a sigmoidal transfer function and depends only on five adjustable coefficients. Model 2 fits the data better but Model 1 has fewer adjustable coefficients. By “adjustable coefficients,” we mean model parameters that are estimated under least squares estimation. Our procedure consists of basically three (3) steps. The first step uses temperature versus time data, at fixed values of r, to obtain the dynamic models for temperature. Next we determine models for the initial temperature and model parameters as a function of r. In the final step we incorporate these models for the parameters, which are functions of r, into T(t,r). We present this work using the following outline: in Section 2 we use the data in Zhao et al. (2006) and develop the dynamic models. Section 3 describes the model building procedures 3 to obtain initial temperature and the dynamic parameters as a function of r. We present and compare both our final models with the three dimensional surface plot in Zhao et al. (2006) in Section 4 and give concluding remarks in Section 5. DYNAMIC MODELING 2.1 Collecting the modeling data In this section we develop dynamic fits for eleven (11) fixed values of r. The data we used come from Figs 5-6 in Zhao et al. (2006) and are given in Fig. 1. The plot on the left gives temperature versus time (t) data at constant values of r and the other one gives temperature versus r data at constant values of t. We used the nine (9) curves in the left plot to obtain T0 for the values of r shown. We estimated these values by mildly extrapolating to t = 0. We added two values of r (80 and 100 mm) using information from the right plot. Their T0 values were obtained from the t = 3 hr curve and their dynamic data were obtained from all the curves at their values of r. Note that the initial temperatures across r represent the temperature profile in the radial direction at t = 0, where we are assuming that t = 0 is the time that the flow of gas was started, i.e., deposition began. Based on the surface temperature plot in Zhao et al (shown later in Fig. 7) we used a maximum temperature of 1000 oC. Next, we use this information to dynamically model temperature at the eleven values of r for Model 1 first and then Model 2. 4 Fig. 1: The plots taken from Zhao et al. (2006)1 used to provide the modeling data. The plot on the left gives temperature versus t curves for different values of r and the one on the right give temperature versus r curves for different values of t. The plot on the left was used to obtain T0 data at each value of r shown except r = 80 mm and r = 100 mm. These two were estimated using the 3 hr curve from the plot on the right. 2.2 Model 1 From the data in Fig. 1, for each of the eleven values of r, we obtained a “best fit” dynamic transfer function for Model 1 using a third order, critically damped transfer function, with a step input. This model form was accepted after evaluating a few low order functions. The Laplace domain form of this transfer function is given by Eq. 1 below: 1000 Tr 0 Tˆr s sˆr s 13 (1) where “^” is used to represent “estimate.” In the time domain, Tr t Tr t Tr 0 (2) where Tr0 is the initial temperature at a distance r from the heater surface and ̂ r is the estimated time constant for the transfer function at a distance r from the heater surface. By applying the method of partial fraction expansion and then transforming to the time domain (see Seborg et al., 2004) and by use of Eq. 2 we get 1 Reprinted from Zhao et al. (2006) with permission from Elsevier 5 t 1000 Tr 0 2 ˆr ˆ ˆ Tr (t ) 1 e r ˆr t t 2 ˆ ˆr t ˆr e r t e 2 T r0 (3) At each value of r, the estimate ̂ r is obtained using the data corresponding to this value of r and nonlinear regression (see Bates and Watts, 1988) to minimize nr SSE r Tr ( t rk ) T̂r ( t rk ) 2 (4) k 1 where nr is the number of samples corresponding to the value of r and trk represents the time of the kth sample at a distance r from the heater surface. SSE is called the “sum of squared errors” and minimization of SSE is called the “Least Squares Criterion,” or “Least Squares Objective Function.” Here, our specific goal is to find the value of ̂ r that satisfies the Least Squares Criterion for each value of r. To achieve this objective we used the “Solver” command in the Microsoft Excel software package. After completion of this step we have a value of ̂ r as well as Tr0 for each r. Therefore, our next step is to produced functional relationships that are dependent on the value of r, that is, ˆr and T0 r . We will do this after completing Step 1 for Model 2. 2.3 Model 2 Dynamically, the advantage of Model 2, the sigmoidal model, over Model 1 is that it allows a steeper approach to the steady state temperature, i.e., the maximum temperature of 1000 o C. Hence, we modeled this process using a special sigmoidal function we developed to have the initial and ultimate response behavior of dynamic transfer functions as shown in Eq. 5 below: 1 1 C 2 r t r 2 1 e T Tr t 1000 Tr 0 r0 1 C where 6 (5) 1 1 C 2 1 e r r 2 (6) 1 C Note that, from Eq. 5, Tr 0 0 Tr 0 Tr 0 and Tr 1000 Tr 0 Tr 0 1000 o C. A 1 C set of adjustable parameters, ˆ r and ˆ r , were estimated using nonlinear regression according to Eq. 4 for each of the eleven values of r. Next we discuss how we obtained the required parameter dependencies on r. r , τˆ r , αˆ r and βˆ r DETERMINING T 0 3.1 T0 r To determine the temperature continuously over all values of r and t, for both Models 1 and 2, continuous functions of T0 r , r , r , and r must be determined. (Note that we are using “(r)” to represent a functional relationship that depends on r and the subscript r to represent the value obtained at one of the eleven values of r using either Eq. 3 for Model 1 or Eq. 5 for Model 2.) We use the results and data from the previous section to accomplish this goal in this section, first for T0 r which is used by both models, then for ˆr for Model 1, and then for r and r for Model 2. A plot of the data for Tr0 against r is given in Fig. 2. As shown, Tr0 monotonically decreases as r increases and reaches a steady state value of 97 C, which must be the temperature at the outer surface of the preform at t = 0. Also, note that the initial temperature at r = 0, T0(0), is shown to be 900 C. This boundary condition was not reported in Zhao et al. (2006) and we estimated it to be 100 C less than the maximum temperature of 1000oC based on Fig. 4 in Zhao et al. (2006), their three dimensional surface plot. We attempted to fit polynomial regression to this data but did not use this form since it could not produce a function with 7 monotonically decreasing behavior and an asymptote at 97 C. Our fitted equation meeting these criteria is r T0 r T0 0 T0 120 e 120 r T0 120 (7) where T0(120) = 97 C and ̂ = 3.143 (determined by nonlinear regression). Note that Eq. 7 has exponentially decaying behavior between r = 0 and r = 120 and the correct limiting behavior (i.e., T0(0) = 900 C and T0(120) = 97 C). As Fig. 2 shows, this equation, for the initial temperature as a function of r, fit the data quite well. 800 Data Fit o T0 ( C) 600 400 200 0 0 20 40 60 80 100 120 Distance from heater surface (mm) Fig. 2: The fit of T0 r using Eq. 7. 3.2 ̂r for Model 1 A plot of the values for ̂ r versus r is given in Fig. 3. In this plot, ̂ r appears to monotonically increase as r increases, and begins to level out at the largest values of r. At r = 0, we set ̂0 = ˆ0 = 4 hr and the maximum value, ˆmax , equal to 98 hr for an optimum fit. Notice that there are no estimated values of the time constant for r < 20 mm. However, one would expect the time constant to continue to decrease monotonically as r approaches zero since the temperature continues to increase in the radial direction towards the heater. In addition, at values of r close to zero, since the dynamic response is fairly fast, model performance should be fairly 8 robust to the choice of r. Nonetheless, in practice, if one desire to more accurately determine the value of the time constant at r = 0, it would be a matter of just collecting the data over time. The fitted equation for the curve given in this figure is: ˆ (r) r ˆ max ˆ 0 ˆ2 ˆ 1 e 2 ˆ r r 2 ˆ r e ˆ r e ˆ ˆ 0 2 (8) where ˆ0 = 4 hr, ˆmax 98 hr, and ˆ 15 .710 mm as determined by nonlinear regression. As shown, Eq. 8 fit these values fairly accurately. Note that Eq. 8 has a form similar to Eq. 3. 120 100 (r) 80 Estimated Values Fitted Curve 60 40 20 0 0 20 40 60 80 100 120 Distance from heater surface (mm) Fig.3: The fit of ̂r for Model 1 using Eq. 8. 3.3 r and r for Model 2 Plots of the values for r and r versus r are given in Fig. 4. In these plots, both r and r monotonically decrease as r increases, and begin to level out at the largest values of r. At r = 0, we set ̂ 0 = ̂0 = 0 hr and ̂ 0 = ̂ 0 = 0.07 hr-1. Similarly, at r = 115, we set ̂115 = ̂115 = -208.972 hr and ̂115 = ̂115 = 0.0217 hr-1 for optimum fits as shown in Fig. 4. As in the case for ̂ r , there are no estimated values for ˆ r and ˆ r for r < 20 mm. However, based on physical considerations, we allowed these values to continue to increase monotonically as r 9 decreased. The fitted equations for both r and r are sigmoidal functions and are given below. 1 1 C 2 r 2 r 115 1 e 1 C (9) where ̂ = -31.3947 mm and ̂ = 0.06433 mm-1, and 1 1 C 2 1 e 2 (10) 1 1 C 2 r 2 r 115 0 1 e 1 C where ̂ = 173.433 mm and ̂ = 0.0315 mm-1, and (11) 1 1 C 2 1 e 2 (12) 0 r (hr) -1 r (hr ) 0.05 0 -210 0 20 40 60 80 100 120 r(mm) 0 20 40 60 r (mm) Fig. 4: The fits of r and r for Model 2 using Eqs. 9 and 11, respectively. 10 80 100 120 t,r PRESENTING T In this section we present our dynamic and spatial models for the temperature field for this process, i.e., Tt,r . The equation for Models 1 and 2 are similar to their dynamic equations (Eqs. 3 and 5) with the subtle differences that the r subscript is no longer necessary since they are valid for any value of r and not just those with measured values and that the initial temperature and the parameters can be determined for any value of r as derived above. Thus, to reflect these properties, the final equation for Model 1 is written as: t t t 1000 Tˆ0 r 2 t 2 ˆ r ˆ ˆ r ˆ r ˆ ˆ Tˆ (t , r ) r 1 e r t e e T r 0 2 ˆr 2 (13) where T0 r is given by Eq. 7 and ˆr is given by Eq. 8. Similarly, for Model 2, 1 1 C 2 ( r ) t ( r ) 2 T 0 (r ) T t , r 1000 T 0 ( r ) 1 e 1 C with 1 1 C 2 ( r ) ( r ) 1 e 2 (14) (15) where r and r are given by Eqs. 9 and 11, respectively. In Figs. 5 and 6 we give the fits of these models to data. Figure 5 gives the temperature versus time (t) responses at constant values of r, in relation to the left plot in Fig. 1 and Fig. 6 gives temperature versus r responses at constant values of t, in relation to the right plot of Fig. 1. Although both plots show excellent fit to the data, Model 2 does the best job in handing the “steep” response behavior. This is best seen by comparing the fits for r = 20 and 30 mm in Fig. 5 and t = 121 and 160 hr in Fig. 6. 11 20 mm 30 mm 35 mm 40 mm 45 mm 50 mm 55 mm 60 mm 80 mm 100 mm 115 mm 20 mm 30 mm 35 mm 40 mm 45 mm 50 mm 55 mm 60 mm 80 mm 100 mm 115 mm 1000 1000 800 800 600 600 400 400 200 200 0 0 0 20 40 60 80 100 Time (hr) 120 140 160 0 20 40 60 80 100 Time (hr) 120 140 160 Fig. 5: Temperature versus time for the eleven values of r for both models corresponding to the left plot in Fig. 1. 1000 1000 3 hr 121hr 40 hr 160 hr 800 40 hr 160 hr 83 hr 0 hr 83 hr 3 hr 121 hr 0 hr 0 hr 83 hr 0 hr 83 hr 800 600 600 400 400 200 200 3 hr 121 hr 3 hr 121 hr 40 hr 160 hr 40 hr 160 hr 0 0 0 20 40 60 80 100 0 120 20 40 60 80 100 120 Distance from heater surface (mm) Distance from the heater surface (mm) Fig. 6: Temperature versus r for six values of t for both models corresponding to the right plot in Fig. 1. Eq. 13 (Model 1) depends only on two adjustable parameters, ̂ through Eq. 7 and ̂ through Eq. 8. Similarly, Eq. 14 (Model 2) depends only on five adjustable parameters, ̂ through Eq. 7, ̂ and ̂ through Eq. 9, and ̂ and ̂ through Eq. 11. With dependence on such a small number of parameters and the excellent fits, the implication is that these model structures capture the phenomenological behavior rather well. This modeling approach falls under the semiempirical class. This class is developed from a combination of data and model structures with phenomenological meaning. This is in contrast to empirical modeling that strongly relies on data and model forms that typically are not physically or biologically interpretable. An advantage of semi-empirical modeling over theoretical or semi-theoretical modeling is that semi-empirical 12 modeling does not require the rigorous development from first principles but is able to obtain forms with phenomenological meaning and model parameters that are often physically interpretable; for example, under Model 1, ̂ r is an estimate of the time constant a distance r from the surface of the heater. (For examples of work using block-oriented modeling, which is a class of semi-empirical modeling that addresses non-linear static and dynamic behavior, see Gomez and Baeyens, 2004; Westwick and Verhaegan, 1996; Shu, 2002; Rollins et al., 2006a; Rollins et al., 2006b; Rollins and Bhandari, 2004; Bhandari and Rollins, 2003; Rollins et al., 2003). In Fig. 7 we show surface plots of temperature over r and t from Zhao et al. (2006) (top plot) and for Models 1 (left plot) and 2 (right plot). As we noted earlier, we set the maximum temperature to 1000 oC in agreement with the figure from Zhao et al. (2006) as shown. As this figures shows, our two surface plots of temperature over r and t compares quite nicely with the top plot from Zhao et al. (2006). Therefore, in summary, we recommend Models 1 and 2 for determining temperature at any value of r and t for the process given in Zhao et al. (2006) and the procedure we have presented in this work for developing models for similar thermal gradient CVI processes. 13 1000 500 0 150 0 100 40 50 15 0 0 0 100 80 40 50 120 80 0 120 Fig. 7: The surface plots for temperature. The top plot is from Zhao et al. (2006)2, the one on the left from Model 1 and the one on the right is from Model 2. With the maximum temperature as 900 oC as indicated in Zhao et al. (2006), the agreement is excellent. CONCLUDING REMARKS In this article we presented a semi-empirical model for the temperature field of a thermal gradient CVI process as a function of space and time. In addition to the model itself, a critical contribution of this work is the illustration of the model building process. This procedure could be used to build these types of models from limited process data and provide critical knowledge 2 Reprinted from Zhao et al. (2006) with permission from Elsevier 14 on the process behavior important to optimal control and optimization. For any real process a much larger study could involve different heater temperatures, gas flow rates, and other boundary and initial conditions. For a fix set of these conditions, as in Zhao et al. (2006), our proposed method offers a way to obtain an accurate knowledge of the thermal field from a set of collected data consisting of even fewer values of r as long as the process is as well behaved over r as it was for the process in Zhao et al. (2006). Our modeling approach is semi-empirical and has several advantages over empirical modeling approaches such are artificial neural networks (ANN). One is that our approach is highly more structured as evidence by the low number of adjustable parameters. To capture the dynamic behavior using an ANN model would require the use of lag variables that carries at least one parameter for each lag variable. Hence, the number of parameters can be enormous depending on the sampling rate and the dynamics. Also adding to the number of variables is the need for one variable for each r value. Thus empirical models can require an extremely large set of parameters that can be quite challenging to estimate under nonlinear regression techniques. Secondly, our proposed approach is a continuous-time one and does not place restrictions on sample data such as constant and equally spaced values as required by empirically-based lag methods. Thirdly, since our approach uses highly structured models with a phenomenological skeletons (i.e., transfer forms), mild extrapolation is less risky using our approach than an empirical model with highly nonlinear behavior and many adjustable parameters. (Note the emphasis we placed on our models having the correct limiting behavior.) Lastly, unlike the empirical model parameters, the model parameters in our approach have phenomenological interpretation. For example, the time constant in Model 1 is τ(r) and in Model 2 it is β(r). 15 The modeling approach that we have presented in this work comes from the area known as system or process identification in the systems engineering literature (Nells, 2001; Gomez and Baeyens, 2004; Westwick and Verhaegan, 1996; Zhu, 2002). Moreover, the work that we have presented in this article has broader impact to the area of systems engineering and transfer function modeling than this application and our recent work in this area (Rollins et al., 2006a; Rollins et al., 2006b; Rollins and Bhandari, 2004; Bhandari and Rollins, 2003; Rollins et al., 2003). To our knowledge, this is the first application to produce a continuous-spatial-temporal model of a real process in analytical form (i.e., closed form) under low parameterization. We were able to do this here because we were able to develop continuous models for the dynamic model parameter and the initial value of the response variable as a function of the spatial variable. Consequently, in other similar type of modeling situations we encourage the use of this approach because the development of analytical continuous-spatial-temporal models offers unique benefits in process understanding, control and optimization. ACKNOWLEDGEMENTS This research was funded by the Future Affordable Multi-utility Vehicle Program DAAD19-03D-0003 under the Department of Defense Appropriations Act administered through Florida A&M University in Tallahassee, Florida. We are also grateful to Ailing Teh for her assistance in the revised manuscript. 16 NOMENCLATURE C CVI nr r SSE t trk T T0 Tr Tr0 The constant given by Eq. 15 Chemical Vapor Infiltration number of samples corresponding to the value of r radial distance from the heater surface Sum of Squares Errors Time time of the kth sample at distance r from the heater surface Temperature initial temperature temperature at distance r from the heater surface initial temperature at distance r from the heater surface Greek letters α adjustable parameter β adjustable parameter γ adjustable parameter λ adjustable parameter τr time constant at distance r from the heater surface τ0 time constant at distance r = 0 from the heater surface Superscript Estimate ^ 17 REFERENCES Bates DM, Watts DG., 1988, Nonlinear regression analysis and its applications (John Wiley and Sons, Inc., New York, USA). Bhandari N, Rollins DK., 2003, A Continuous-Time MIMO Wiener Modeling Method, Industrial and Engineering Chemistry Research, 42(22):5583-5595. Birakayala N, Edward EA., 2002, A reduced reaction model for carbon CVD/CVI processes, Carbon, 40(5):675-683. Delhaes P., 2002, Chemical vapor deposition and infiltration processes of carbon materials, Carbon, 40(5):641-657. Golecki I, Morris RC, Narasimhan D, Clements N., 1995, Rapid densification of porous carbon/carbon composites by thermal-gradient chemical vapor infiltration, Appl Phys Lett , 66(18):2334-2336. Gomez JC, Baeyens E., 2004, Identification of Block-oriented Nonlinear Systems using Orthonormal Bases, J of Process Control , 14:685-697. Jiang K, Li H, Wang M., 2002, The numerical simulation of thermal-gradient CVI process on positive pressure condition, Mats Letts, 54:419-423. Li H, Li A, Bai R, Li K., 2005, Numerical simulation of chemical vapor infiltration of propylene into C/C composites with reduced multistep kinetic models, Carbon, 43(14):2937-2950. Nells O., 2001, Nonlinear System Identification, (Springer, Germany). Probst KJ, Besmann TM, Stinton DP, Lowden RA, Anderson TJ, Starr TL., 1999, Recent advances in forced-flow, thermal-gradient CVI for refractory composites, Surf Coat Technol, 120-121:250-258. Rollins DK, Bhandari N., 2004, Constrained MIMO Dynamic Discrete-Time Modeling Exploiting Optimal Experimental Design, J of Process Control, 14(6):671-683. Rollins DK, Bhandari N, Bassily AM, Colver GM, Chin S., 2003, A ContinuousTime Nonlinear Dynamic Predictive Modeling Method For Hammerstein Processes, Industrial and Engineering Chemistry Research, 42(4):861-872. Rollins DK, Bhandari N, Chin S, Junge TM, Roosa KM., 2006, Optimal Deterministic Transfer Function Modeling In the Presence of Serially Correlated Noise, Chem Eng Research and Design, 84(A1):9-21. Rollins DK, Bhandari N, Hulting S., 2006, Continuous-time Block-oriented Predictive 18 Modeling of the Human Thermoregulatory System, Chem Eng Sci, 61:1516-1527. Seborg DE, Edgar TF, Duncan AM., 2004, Process Dynamics and Control (John Wiley and Sons, Inc., Danvers, USA), pp. 59-65. Tang Z-h, Qu D-n, Xiong J, Zou Z., 2003, Effects of infiltration conditions by a directional-flow thermal gradient CVI process, Carbon, 41(14):2703-2710. Vignoles GL, Langlais F, Descamps C, Mouchon A, Le Poche H, Reuge N, Bertrand N., 2004, CVD and CVI of pyrocarbon form various precursors, Surf and Coat Tech, 188-189:241-249. Westwick D, Verhaegan M., 1996, Identifying MIMO Wiener Systems using Subspace Model, Identification Methods, 52:235-258. Zhang WG, Hüttinger KJ., 2003, Densification of a 2D carbon perform by isothermal, isobaric CVI: Kinetics and carbon microstructure, Carbon, 41(12):2325-2337. Zhao JG, Li KZ, Li HJ, Wang C., 2006, The influence of thermal gradient on pyrocarbon deposition in carbon/carbon composites during CVI process, Carbon, 44(4):786-791. Zhu Y., 2002, Estimation of an N-L-N Hammerstein-Wiener Model. Automatica, 38:1607-1614. 19