To better understand biomass pyrolysis, the three main components of biomass—cellulose, hemicellulose and lignin—are investigated separately. For this purpose, extracted and purified components have been purchased from commercial suppliers. Cellulose (product number: 22182, Sigma Aldrich) was produced from spruce by acid washing with a final purity of >99.99%. Xylan from beechwood (purity >95%) was used as a model compound for hemicellulose (Product code: P-XYLNBE, Megazyme Ltd.). Alkali lignin was purchased from Sigma Aldrich (product number: 370959).

The pulverized materials were classified into different mesh size fractions, whereof the size fraction 90–125 µm was used for the pyrolysis experiments in the FBR. Due to the limited availability of sample material for xylan, only the size fraction 60–90 µm could be used. Table 1 provides information from ultimate, proximate and microscopic particle size analysis for all three components.

The pyrolysis kinetics of lignocellulosic components for low temperature and heating rate conditions are determined by thermogravimetric analyses (TGA) with a TGA8000 from PerkinElmer Inc. Pulverized particles with the size below 75 µm are used to minimize the deviation from the intrinsic kinetics due to the limitation in heat conduction. Around 0.8–1.2 mg of sample mass is loaded and spread at the bottom of an alumina crucible (7 mm diameter and 2 mm height) as a thin layer to minimize the effect of bulk diffusion. The reaction analyzer is purged with N₂ (purity >99.996%) in the vertical direction down to the crucible through the balance chamber. The gas flow rate is kept constant at 50 ml min⁻¹ at standard conditions (25°C and 105 kPa). The sample is heated from 303 to 1173 K with a constant heating rate of 5 K min⁻¹.

In the fluidized bed reactor (FBR), the pyrolysis kinetics are derived from the time-resolved volatile release rate. The release rate is calculated from a mass balance around the system. For this balance, the gas inflow is regulated and the exhaust gas species concentrations are continuously measured using a Fourier-transform infrared spectrometer. The FBR combines high particle heating rates around 10⁴ K s⁻¹ with moderate temperatures in the range from T _FB = 573–973 K.

A sketch of the reactor system can be found in Figure 1. The reactor itself consists of two coaxial, vertically mounted ceramic pipes positioned in an electrically heated furnace. The fluidized bed, representing the reaction zone at position ①, is generated by a nitrogen gas flow passing through a porous sintered glass plate. The bed is formed by inert Al₂O₃ particles in the size range of 112–180 µm. The empty volume above the bed is blocked by a ceramic filling to suppress diffusion processes and to guarantee a quick transport of all gases from the reaction zone to the gas analyzer. The reactor temperature is regulated through the electric furnace and is continuously monitored with a thermocouple immersed in the bed, to ensure constant and homogeneous reaction conditions.

Small amounts of fuel (15–50 mg) are injected batch-wise into the fluidized bed from the top. To ensure a consistent and controlled atmosphere in the reactor, the fuel is introduced through a double-lock system (position ⓪), which is purged at each time with gas from the fluidized bed before injection. To enhance the transport of all fuel particles from the lock to the reaction zone, a gas flush of 40 ml N₂ is used. This reduces the sticking of particles to the feed pipe walls and reduces residence times under non-isothermal boundary conditions during the feed process. Particles arrive in the bed approximately 150 ms after feeding has been initiated and instantly start to devolatize. Released gaseous reaction products mix with the fluidizing gas and the gas mixture leaves the reactor through the exhaust gas pipe.

The full gas flow is analyzed with an Agilent Cary 670 Fourier-transform infrared (FTIR) spectrometer at position ②. The FTIR is equipped with a PIKE Technologies gas cell with 2.4 m optical path length and 100 ml volume. The cell and the exhaust gas pipe connecting it to the reactor are heated to a temperature of 453 K to prevent tar and water vapor condensation within the equipment. The infrared spectra are recorded with a frequency of 10 Hz and a spectral resolution of 4 cm⁻¹ before being analyzed in the wavenumber region from 600 to 6500 cm⁻¹ using multivariate regression analysis. The regression analysis relies on measured reference spectra of single species with known volume fractions to determine the pyrolysis gas composition φ _i(t) for 22 different gas species (Pielsticker et al., 2017).

Based on the measured volume fractions and a mass balance around the reactor system, the volatile release rate can be derived with Equation 1, which on the right hand side is structured in three terms: d y vol, exp ( t ) d t = ∑ i = 1 22 ( ρ N 2 , n ⋅ M i M N 2 ⋅ V ̇ N 2 , n m 0 ⋅ φ i ( t ) 1 − ∑ j = 1 22 φ j ( t ) ) (1) The first term contains constants (N₂ density at standard conditions ρ N 2 , n , molar mass M) which are identical for all experiments. The quantities given in the second term (volume flow V ̇ N 2 , n , sample mass m ₀) are parameters of the respective test series, which are adjusted to temperature and fuel reactivity but remain constant within an individual experiment. The third term contains the time-dependent volume fractions φ(t) measured individually for every single experiment.

Further details on the reactor setup, its implicit limits and the execution of experiments (Pielsticker et al., 2019a), the evaluation routine (Gövert et al., 2017; Pielsticker et al., 2019a), as well as the imposed limits of the gas transport on the detectable reaction rates (Pielsticker et al., 2018a; Pielsticker et al., 2019a), can be found in the given references.

Two empirical models of different complexity are used to approximate the pyrolysis process of biomass and its basic components: The single first-order reaction (SFOR) model and the two-step model with intermediate formation (2-step). As the decomposition of biomass is a very complex process with more than several hundred reactions and species being involved, both empirical models are not able to describe the pyrolysis process in detail. Nonetheless, the models can adequately approximate the volatile release characteristics as shown in previous studies (Ontyd et al., 2020; Pielsticker et al., 2018b; Pielsticker et al., 2019a). Both models can be calibrated using data from the ex-situ gas analysis. Model types that require information regarding the time-dependent change of the solid phase are not applicable for the FBR setup as char sampling during the pyrolysis process is not possible. Empirical pyrolysis models with a limited number of equations bring the advantage of low numerical effort and thus are frequently used in large-scale CFD simulations (e.g. Nicolai et al. (2020)).

Figure 3 shows the final volatile yield y _vol,∞,i(T) of the three components cellulose, hemicellulose and lignin at different temperatures determined by integrating Equation 1 over the full range t ∈ [0, ∞]. In the practical application, the experiment is stopped when the measured concentrations fall below a certain threshold. While for cellulose and hemicellulose, the gas concentrations level out around zero, for lignin the concentrations of H₂O, CO₂, and CO stay constantly above zero indicating that decomposition reactions still take place. As a consequence, the calculated “final yield” can strongly depend on the integration time and should thus be treated with caution. This effect is also visible in the peak yield shown in Figure 3: the mass fraction of the primary reaction peak (y _vol,exp,peak) sometimes accounts for less than 50% of the total mass released.

For all three components, the total yield of volatile products increases with temperature, showing a degressive characteristic approaching the volatile yield of the proximate analysis. Typically, yields from fast pyrolysis are slightly above the volatile content from the proximate analysis. Thus, a small increase is expected for temperatures above the current range. The final mass fractions experimentally determined in this work are within the scatter band of measured data available in the literature for comparatively fast pyrolysis ¹ of cellulose, hemicellulose and lignin.

For modeling the temperature-dependent final volatile release, a modified logistic distribution function y vol , m o d , ∞ , i ( T ) = C 1 1 + C 2 ⋅ e − C 3 ⋅ T − C 4 C 5 − 1 (8) is used, whose coefficients C ₁–C ₅ are determined by minimizing the squared errors between the model and the experimental data (own experiments and literature) and are summarized in Table 2. Using also experimental data from literature—especially in the low-temperature region—is necessary to obtain stable fit results. The corresponding curve in Figure 3 shows that the model can adequately represent the temperature dependence of the final volatile release (R ² = 0.892, 0.832, and 0.946 for cellulose, hemicellulose and lignin) and that the model also allows stable extrapolations for higher and lower temperatures.

The mass fractions of the analyzed light gas species are summarized in Figure 4. As evident from the figure, H₂O, CO₂, and CO constitute the largest yield fractions for all three fuels at all temperatures. This is not surprising, considering that (hemi-)cellulose comprises around 50 wt% oxygen and lignin about 30 wt% (cmp. fuel analysis in Table 1). Hence, many of the complex hydrocarbon molecules cracked during pyrolysis are instantly attacked by oxygen radicals also formed in this process and thereby oxidized. The absolute yields of H₂O and CO₂ (not shown in Figure 4) fluctuate between experiments and operating points but show no clear trend with temperature. As the absolute yields of most other gases, including CO and CH₄, increase with temperature, the combined yield fractions of H₂O and CO₂ show a slow and steady decline over T.

The absolute and relative yields of CH₄ are similar for cellulose and lignin, while hemicellulose releases less CH₄ at all temperatures and especially below 873 K. The combined yield fraction of other gases—mostly short-chained hydrocarbons—is highest for hemicellulose, followed by cellulose and then lignin. For the latter two, the fractional and the absolute yield of other gases increase with temperature, while they stagnate for hemicellulose.

Comparing the results presented in Figure 4 with literature is not trivial. The available literature revolving around “fast pyrolysis” is using very different experimental setups, ranging from tube furnaces (Lv et al., 2013) over fixed beds and Py-GC/MS devices (Zhao et al., 2018) to TGAs and film or wire mesh heaters (Hoekstra et al., 2012; Ansari et al., 2019). Heating rates in those experimental setups vary over a wide range and are often significantly lower than those achieved in the FBR used in this study—ranging from 1 to 100 K min⁻¹ in TGAs over estimated 10000 K min⁻¹ in tube furnaces (Lou and Wu, 2011) to as high as 7000 K s⁻¹ on a wire mesh (Hoekstra et al., 2012). Data on water and hydrogen yields are often missing in the reported datasets.

Nonetheless, global trends are often comparable: Lv et al. (2013) report the same order for CH₄ yields at 873 K and above, where y _CH4,Lig > y _CH4,Cel > y _CH4,Hem. Lv et al. (2013) and Zhou et al. (2014) equally find, that the CO yield is highest for cellulose. Lv et al. (2013) further reports that the yields of short-chained hydrocarbons are small below 873 K but continuously increase with temperature, as found in this work. Concerning CO₂ yields, there is no consensus amongst the sources listed in this chapter: some report continuous increases with temperature (Scott et al., 1988; Lv et al., 2013), others report CO₂ yield maxima around 773–873 K for one or multiple of the investigated biomass components (Lou and Wu, 2011; Lv et al., 2013). It is unclear whether these differences are attributable to different primary decomposition steps or an effect of secondary gas-phase reactions.

As outlined in section 2.4, the single first-order reaction model has one adjustable parameter, the characteristic reaction rate r _SFOR at a given reaction temperature. If this dependency is modeled using an Arrhenius approach (Equation 4), the modeling is based on two parameters, the pre-exponential factor A and the activation energy E _a. The following two sections illustrate how those kinetic parameters are determined from the experimental results obtained in the TGA and FBR setups.

As described in section 2.4.2, the 2-step model has three adjustable parameters, the two reaction rates of the primary (r _prim) and secondary (r _sec) decomposition as well as the share of intermediate formed (ϕ _int). Similar to the SFOR model, these parameters are determined from single experiments at fixed fluidized bed temperatures via a least-squares fit algorithm. The additional fit parameters lead to a better overall approximation of the volatile release rate of cellulose and hemicellulose ( R ̄ cel 2 = 0.935, R ̄ hem 2 = 0.955) compared to the SFOR model ( R ̄ cel 2 = 0.795, R ̄ hem 2 = 0.894) experiments with a sharp peak in the beginning followed by a slow decrease in the late phase of the experiment (e.g. the cellulose peak shown in Figure 6) are captured significantly better. Lignin benefits the most from changing the model: While the SFOR model ( R ̄ lig 2 = 0.442) is mainly unable to adequately reflect the pyrolysis behavior, the 2-step model ( R ̄ lig 2 = 0.959) offers a fit quality similar to those of cellulose and hemicellulose.

Figure 9 shows the temperature dependency of the primary and secondary reaction rate in an Arrhenius diagram (Figure 9A) and the intermediate fraction (Figure 9B). The Arrhenius diagram reveals two distinguished reaction rates with different orders of magnitude for each component within the investigated temperature range. Similar to the SFOR model, a transition between two reaction regimes is visible for hemicellulose while cellulose and lignin appear to react in the same reaction regime over the entire range of investigated temperatures in the FBR. The calculated Arrhenius parameters are listed in Table 4.

The intermediate fraction shown in Figure 9B shows similar trends for all three components: the lower the temperature, the higher the fraction of intermediate formed. Consequently, the overall release rate is more strongly controlled by the secondary reaction step for low temperatures, while the first reaction is the deciding factor for higher temperatures. For cellulose and hemicellulose, the influence of the secondary reaction even vanishes completely above 873 K. In the case of negligible intermediate fractions also values of the secondary reaction rate become arbitrary and are not considered in the Arrhenius fit. The distributions of ϕ _int(T) are modeled with a double exponential function: ϕ int, i ( T ) = ϕ int,LT,i − ϕ int,LT,i − ϕ int,HT,i ⋅ 1 − exp − A ϕ , i ⋅ e E a, ϕ , i R ⋅ T (11) This approach is comparable to the description used by Nunn et al. (1985) to approximate the temperature-dependent mass release in their study and has already been proven to be suitable to model the intermediate fraction (Pielsticker et al., 2019b). All parameters for the modeling of ϕ _int,i(T) are also given in Table 4.

To evaluate the overall model’s accuracy, a suitable quality criterion must be defined. The investigation of different criteria has identified the weighted, normalized absolute error W * = ∑ i d y vol, exp, i d t − d y vol, mod, i d t 2 ⋅ d y vol, exp, i d t d y vol, exp d t max 2 (12) as the most meaningful quality measure. It simultaneously is highly sensitive to the maximum reactivity (highest intensity of release rate) while being robust to non-reaction-related influences like measurement frequency, recorded duration of the experiment or signal intensity.

The overall model performance is calculated using the final yield model in combination with both reactivity models. Figure 10 shows the calculated values of the quality criterion in logarithmic space. To make the numerical values tangible, reference values are determined using the experimental standard deviation σ and multiples of it. The resulting ranges of log(W*) are given in Figure 10 as areas of different grayscale. Desirable is an overall model performance within the ±2σ range corresponding to values of log(W*) < −0.75.

For cellulose and hemicellulose, most of the points fulfill the aspired criterion log(W*) < −0.75 and all model predictions fall within the ±3σ range. The two highest temperatures for cellulose and the highest one for hemicellulose have been discarded from the analysis. At these conditions, the chemical reaction and the enlarged gas volume flow due to the gas flush during fuel injection act on the same time scale. Since the convolution function only works with a constant volume flow, it is impossible to reflect the effect of an unsteady volume flow during the flush. Cellulose reveals a slightly better prediction with the 2-step model (log(W*) = −1.09) compared to the SFOR model (log(W*) = −0.93) and an overall slightly worse approximation than hemicellulose. For hemicellulose, no significant difference regarding the model selection can be obtained (SFOR: −1.26, 2-step: −1.26). For lignin, the largest difference between both model approaches is visible: While the SFOR model—especially at lower temperatures—gives the worst prediction (log(W*) = −0.71), the 2-step model (log(W*) = −1.46) is well suited to describe the flash pyrolysis process of lignin.

By comparing the prediction behavior of the SFOR model with the experimental results for lignin in detail (Figure 11), it turned out that the SFOR model underpredicts the conversion in the beginning and overpredicts the release rate in the late phase of the reaction. Two modifications to the original (overall) SFOR model formulations are investigated to analyze whether an improvement can be obtained and how these improvements compete with the 2-step prediction:

The first modification confines the region of interest to the primary peak and thereby avoids the overpredicted late phase. Instead of the final volatile yield, only the mass fraction released in the primary peak region (cmp. triangle symbols in Figure 3) is considered in Equation 2. The kinetic parameters have been adjusted accordingly (Figure 7 and Table 3). The modification allows for a good representation of the peak reaction behavior but predicts a completed reaction shortly after the peak. Consequently, this approach reflects the reaction behavior of only parts of the lignin. Nevertheless, this approach might be a suitable compromise if only short residence times are of interest (e.g. pulverized fuel combustion).

The second modification implements a logic that reduces the reaction rate depending on the reaction progress. The idea of this approach is equivalent to the assumption that the reacting material consists of many different partial masses, which react individually with an individual reaction rate. Typically the different reaction rates are expressed via a distribution function. In this case, a Gaussian normal distribution—with mean value μ _r and standard deviation σ _r as fitting parameters—is used to approximate r(y): y vol ( t ) y vol , ∞ ( T ) = 1 σ r ⋅ 2 π ∫ − ∞ r ( t ) − exp − r ( t ) − μ r 2 2 ⋅ σ r 2 d r (13) The Matlab built-in function for inversion is used to determine r(y). Figure 11 shows, that this model approach is also much more suitable to predict the volatile release compared to the standard SFOR model. It has the advantage that it can correctly predict both, the peak reactivity and the final volatile yield. Nevertheless, in comparison with the prediction of the 2-step model, larger deviations to the measured conversion rates in the late phase of the experiment are revealed. Despite this, the SFOR model with distributed reactivity represents a good alternative to the 2-step model. This proposed approach is comparable to the frequently used distributed activation energy model (DAEM) so far only calibrated with TGA data (Cai et al., 2014). In future works, it needs to be investigated to what extent mean value μ _r and standard deviation σ _r of the reaction rate show temperature-dependent behaviors, whether these can be modeled and a determination of DAEM kinetic parameters (mean value and standard deviation of the activation energy) is possible.