The first aim of the work is to assess the robustness of the new reduced cost and small form factor MEMS FPI NIR sensor over a 9 month period during which 14 batches of placebo granules were manufactured and dried in a fluidised bed (Table I). The objective was to repeat essentially identical batches in order to evaluate the robustness and consistency of the sensor, and the predictions based on models derived from the NIR spectra. The system was however subject to changes uncontrolled variables: such as batch to batch variations in granulate (e.g. size, moisture content, storage time) and the ambient and air inlet temperature and humidity (experiments started with bx1 in August to bx 14 in mid December). To probe the sensitivity of the PLS moisture predictions to a change in flowrate we reduced the air flowrate from 850 to 600 L/min in the last 4 batches.
Overall, the MEMS-FPI sensor performance showed a very satisfactory stability and reproducibility. The spectral signal remained stable and repeatable under all ambient conditions. The bright reference intensity levels and spectra shape measured at the beginning of each experiment remained similar. The device proved to be free of interferences generated by mechanical vibrations; e.g. spectral readings did not alter when the device was installed directly next to the fluidised bed vessel. Ambient temperature variation did lead to minor variations. However, using a second spectrometer it was demonstrated that this was in fact due to effects of temperature on the transmission through the fibre optic cables, rather than on the light source or the MEMS-FPI sensor.
In the fluidised bed process granules are dried using ambient air of uncontrolled humidity. As a result the drying rate varies from batch to batch, with the final drying end-point determined by relative humidity conditions found on the day when a batch is dried. Table I summarises the observations for all batches. A typical data set for a drying experiment is shown in Fig. 3: Granules are charged cold and fluidised by a gas stream of 600 L/min at ~25 ° C. As water evaporates, heat is removed from the gas stream resulting in a ~10 ° C temperature difference between the bed and the fluidising gas. The moisture content measurement is based on a PLS model (Eq. 1) based on 6 reference batches and validated using the NIR spectra from a further 8 validation batches. Moisture content is monitored continuously by NIR using the PLS model, and measured once every ~10 min by sampling and performing offline analysis by LOD. When the granule water content is reduced by ~85% the evaporation rate drops and the bed temperature rises. Equilibrium between the inlet gas (ambient RH) and the granules (fw≈1.5–2 wt%) is reached after ~90 min at 600 L/min.
To test the robustness of the NIR measurement and the associated PLS model, drying was performed using two different flow rates (600 and 850 L/min) at constant air inlet temperature (25 ± 3°C). As expected, the flowrate strongly affects the drying rate: thus, the drying time reduces to ~60 min when the flowrate is set at 850 L/min. Even though the PLS model is based on data from runs 1 to 6, all operated at 850 L/min, it still correctly predicts the moisture content of runs at 600 L/min without further chemometric analysis, since only the drying rate, but not the sample composition, is changed. Thus, the fitted parameters for Eq. (1) successfully correlate fw to the NIR spectra for the reference batches, and the same parameters allow prediction of moisture content for all validation batches irrespective of the flowrate. This demonstrates (i) the robustness of the data strategy applied (e.g. data treatment, data treatment prior to PLS analysis) and (ii) the excellent stability and robustness of the MEMS-FPI NIR sensor.
When drying from ~33 to 10 wt% moisture contents it was observed that the granules shrank due to the loss of 26% of their original mass. At this stage no significant fines were observed. At lower moisture levels (5–1%), granules appear dried at the surface. Granule size continue to reduce, but now by attrition which generated a significant quantity of fines that were observed to build up on the fluidised bed filters, resulting in increasing pressure drop across the outlet filters. The size reduction, coupled to the reduction in moisture content did appear to increase the intensity of the reflected signal as the drying process progressed, but this did not seem to have a significant impact on the moisture content predictions.
A data analysis strategy was designed (Fig. 4) to demonstrate that the low cost NIR sensors delivers high quality data robustly during fluid bed drying on placebo pharmaceutical granules; a typical processing scenario with a very low signal to noise ratio. We tested three applications: Moisture monitoring, end-point detection and process analysis (mass transfer monitoring) that will be discussed in more detail below
NIR Signal Acquisition during Fluidised Bed Drying
The presence of large fluctuations in individual NIR scans when sensing in situ from a fluidised bed dryer has been previously reported and necessitates intensive data pre-processing steps and/or modification of the way a spectrum is measured (e.g. scoop devices to hold the sample in place while measuring) (28). The contact between the fluidised material and the tip of the NIR probe is variable and characterised by air gaps appearing in front of the sensing area at random intervals. This produces abrupt changes in the spectra collected from scan to scan. Figure 5a, b, c shows 10 consecutive single spectral readings and their corresponding average (darkened line) for three different time periods of the drying process (a, t = 0 min; b, t = 41.2 min; and c, t = 81.2 min). In this work 10 single spectral scans were averaged to give a single NIR spectrum every 10 s. This procedure is exemplified in Fig. 5, where the average of the 10 single scans was converted to the absorbance spectra. The resulting absorbance signal still shows a relatively high level of noise, which is further reduced with a moving averaging filter that averages the data of the currently aquired spectrum with the previous 74. The large filter window (750 s) provides a smoother variation of the spectral observations over time and only slightly reduces the response time of the moisture content predictions.
After the pre-processing steps, the drying evolution can be clearly observed in the spectra (Fig. 6): the absorbance of the -OH absorption band (1940 nm), associated with the presence of water in the material, strongly reduces as drying progresses.
Fouling is another issue observed when acquiring in situ NIR spectra in a fluidised bed system. This problem often occurred at the beginning of the drying process, since granules with a moisture content above 20 wt% are very wet, and have a strong tendency to stick to the probe’s window. This causes more reflection, and thus higher intensity NIR spectra, which shifts the signal to greater values compared to normal operation with a clean window. Fortunately, continuous collision of granules on the probe window causes a degree of self-cleaning. Hence, if granules stick for periods significantly smaller then 750 s, the pre-processing steps will reduces the impact of fouling on moisture content measurements.
Moisture Content Prediction with the NIR Sensor
A PLS model was developed by relating the pre-processed NIR spectra from the initial six batches (Table I), to the offline measured LOD moisture content. For the six calibration batches (Bx1 to 6), different, narrow spectral regions including the spectral range corresponding to the strong moisture band (1900–2000 nm) were tested separately but the smallest overall moisture prediction error was found using the full spectral range available from the sensor (1750–2150 nm).
Figure 7 shows the results obtained for six selected batches comparing the LOD analytical moisture content (circles) with the predicted moisture profiles obtained from the online NIR spectra and the PLS regression model (continuous line). The figure shows (i) two calibration batches used for developing the PLS model (batches 2 and 4, using 850 L/min), (ii) two validation batches with the same flow rate (batches 7 and 9, using 850 L/min), and (iii) two validation batches using a slower flow rate (batches 13 and 14 using 600 L/min). Similar plots showing the results obtained for all fourteen batches can be found in the Supplementary Material. Generally, the moisture content resulting from the NIR spectra provides a good estimate of the data measured offline by LOD when the moisture content fw < 20wt%.
A direct comparison of the moisture content measured using LOD with the predicted results from online NIR spectra is given in Fig. 8. As the PLS model fitted log10fw, we expect the relative error in the prediction to be constant for different levels of water content.
The average residual for all N LOD measurements (~200 LOD values) may be defined as:
$${e}_{rel}=\sqrt{\frac{1}{N}\sum \limits_{i=1\ to\ N}{\left(\frac{f_{w. NIR}-{f}_{w. LOD}}{f_{w. LOD}}\right)}_i^2}=13\%$$
(6)
From Fig. 8 it is clear that the error in the NIR based water content is independent of the absolute extent of the moisture content. The error for the calibration batches is ± 10%. However, all models over predict the LOD concentration below 5 wt%: 1% higher for the calibration batch, 18% for the 850 L/min and 8% for the 600 L/min validation batches. The high deviation at low water levels for the 850 L/min batches may be due to an uncontrolled parameter such as the humidity.
Comparing results to previous studies is difficult as the water concentration ranges used to build principal component models vary significantly. The pre-processing strategy was similar in all cited cases, averaging a large number of scans (from 32 to 300 vs 750 in this study) to compensate for the spectral noise. Peinado et al. (23) reported data for water contents fw between 0.6 wt% to 2.8 wt% with a relative error εrel ≈ 15% using an ABB Fourier Transform Process Analyser Near Infrared spectrometer, with thermo electrically cooled InGaAs detectors (ABB-FTPA2000260). Fonteyne et al. (19) reported data for fw in the range of 3.5 to 7 wt% with a relative error εrel = 5% using a Matrix™ –F Duplex, Bruker Optics Ltd., FT-NIR spectrometer (32 single scans averaged; using 1000–2220 nm for analysis). The residuals obtained using a commercial dispersive spectrometer varied ±4 wt% over 4-20 wt% water content (32 single scans averaged; using 1100–2500 nm for analysis) (28). The results obtained for the calibration and the validation batches with the novel MEMS-FPI sensor (εrel=13%) were similar to those reported with conventional spectrometers. We did however observe significantly larger relative errors below 5 wt% water in the validation batches. The absolute errors remained low (~0.4 wt%). Based on the operation over a significant period, we feel this is more likely to be due to changes in uncontrolled parameters
Process End-Point Detection from MSPC Charts
A different application to evaluate the performance of the MEMS-FPI is the detection of the process end-point from the NIR signal. We use the MSPC model (Eq. 2) based on the end-point spectra from the calibration batches. This model required two principal components, PC1 marks the inverted water band and PC2, with a less interpretable shape, is required for the description of batch-to-batch variability (the PC loadings and related description are provided in the supplementary information S4). Qstat control charts were calculated from Eqs. 2–4 for validation batches by projecting NIR spectral observations (using same pre-processing procedure as before) onto the developed model. Figure 9 shows the Qstat MSPC charts obtained for six validation batches. Detected end-points are indicated with a yellow diamond marker in the Qstat control charts. Batches 1, 5, 8 and 14 reached a final moisture content below 2% (on-specification), batches 12 and 13 did not (off-specification). For comparison, moisture content levels from the NIR spectra were compared in the MSPC control charts for batches 5 and 13 (moisture axis is at the right of the plot). These plots in Fig. 9 show that based on moisture levels the endpoint would have been delayed for batch 5; both Qstat and the moisture level agree that batch 13 is off specification.
The results in Fig. 9 also confirm that a lower gas flow rate (batches12–14, see Table I) significantly increases the end-point process time (compared with on-specification batches, 1, 5 and 8). Similar results were observed for the other validation batches (see Supplementary Material). The spectral quality of the NIR sensor and its robustness is clearly sufficient for the construction and application of PCA-based MSPC models. The device has clear potential in end-point detection applications, promising a significant reduction in offline moisture measurements.
Process Monitoring (Mass Transfer Resistance)
The promise of low cost NIR devices lies in the common availability of online compositional analysis. Our statistical analysis (PLS and MSPC) demonstrates the low cost MEMS-FPI NIR sensor to be a device that is suitable for composition measurement, with sufficient performance compared to conventional sensors when applied to fluid bed systems. To date NIR data has not been used in conjunction with mechanistic drying models to provide scale up data. As can be seen in Fig. 4, such analysis is complex, and requires the fusion of data from multiple sensor, as well as an understanding of material properties such as water adsorption isotherms and the vapour pressure of water. Methods to determine the bed moisture content and drying rate from temperature and humidity data show a significant deviation from samples analysed by LOD (33). To demonstrate the utility and value of continuous composition data in process monitoring we developed a methodology to monitor the process by evaluating the mass transfer resistance(s) from the available data. These resistances underpin fluidised bed scale up calculations. An overview of the mass transfer model is given in Fig. 10.
Step 1: Obtain the Molar Drying Rate
To evaluate the mass transfer we first must convert the Drying curve fw(t), to the drying rate in mols/s by differentiation of the water content \({N}_w=\frac{1}{0.018}\ \frac{m_s\ {f}_w(t)}{1-{f}_w(t)\ }\) in the bed. The slope of the dying curve results from linear regression of a line to 36 data points either side of time t (a total of 72 points over 720 s). The standard error in the slope obtained is between 3%–5%. An example of fw, and its derivative dfw/dt are shown in Fig. 11 row 1. The first graph shows the conventional drying curve consisting of the initial transient phase followed by the constant drying rate period (20–60 min, \(\frac{d{f}_w}{dt}\approx 0.8\ wt\%/\mathit{\min}\)) and the falling rate period and finally equilibrium (fw ≈ 1.6 wt % ). Row 2 of the same figure shows the molar drying rate at approximately 0.15 mol/s in the constant drying rate regime.
Step 2 Work out the Driving Force
Granules consist of solids bound together by a binder fluid (Fig. 10a). Water is present in liquid bridges, but also adsorbed to some of the solid materials. The process scheme (Fig. 10b) shows the location of water in different environments (e.g. different phases and segregated domains) with arrows representing mass transfer between the environments. The mass transfer process is a diffusion process driven by the concentration gradient between water at the source in the granule and the water in the fluidising gas (the sink). As the liquid and solid phases in the granule are in intimate contact we assume equilibrium hence the gas phase concentration over (in equilibrium with) the solid (\({C}_S^{\ast }\)) and liquid (\({C}_L^{\ast }\)) are the same. As at some point the liquid phase will disappear, we focus our attention on obtaining \({C}_S^{\ast }\) as function of temperature and water content. Thermodynamically, the vapour pressure over a solid or liquid phase is represented by the product of the activity of water in that phase aw and the vapour pressure of water at the bed temperature, \({P}_w^{*}\left({T}_{bed}\right)\):
$${C}_S^{\ast }=\frac{a_w{P}_w^{*}\left({T}_{bed}\right)}{R\ {T}_{bed}}\ \frac{mol}{m^3}$$
(7)
The equilibrium vapour pressure of pure water is given by the Antoine equation (34), supplementary material), here expressed relative to the vapour pressure at a reference temperature of 20 ° C:
$${P}_w^{*}\left(T\left({}^{\circ}C\right)\right)=2339.1\ast {e}^{-4078.8\ast \left(\frac{1}{236.63+T\left({}^{\circ}C\right)}-\frac{1}{256.63}\right)}\pm 1\ Pa$$
(8)
The water activity (aw) follows from the water adsorption isotherms of the materials present in the placebo.Footnote 2 Such isotherms relate aw to the water content on a dry weight bases (Fig. 12a). The GAB correlation, an extended BET equation developed by Guggenheim, Andersen and de Boer (35,36), expresses water content of solid i\(\left({F}_{w_i},\mathrm{dry}\ \mathrm{basis}\right)\) as function of the water activity (Table II, and supplementary material):
$${F}_{w_i}\left({a}_w\right)=\frac{m_{w_i}}{m_{s_i}}=\frac{{k_w}_i{a}_w{C}_{GA{B}_i}{m}_{o_i}}{\left(1-{k}_{w_i}{a}_w\right)\left(1+\left({C}_{GA{B}_i}-1\right){k}_{w_i}{a}_w\right)}$$
(9)
Table II Fitting Parameters for the GAB Equation of the Materials in the Placebo Formulation At the beginning of the drying the Hypromellose 2910 (5 wt% dry basis) and Ac-Di-Sol (1.5 wt%) contain up to 50% of the adsorbed water with the remainder adsorbed onto the Avicel (29 wt%). As the granule dries, this reduces to ~10% and most of the remaining water is associated with the Avicel. Assuming the materials do not interact, an aggregate isotherm may be constructed by combining the water content adsorbed by the various materials at the same water activity (Fig. 12b). The aggregate isotherm correlates the granule’s water activity to Fw = mw/madsorb, where madsorb is the combined mass of Avicel, Hypromellose and Ac-Di-Sol (Fig. 12c, fitted GAB parameters in Table 22).
Step 3: Link the Molar Drying Rate and the Driving Force
The mass transfer process can be represented with a resistance model ((41), Fig. 10c) that sees a “current” of water (\(MT{R}_w,\frac{mol}{s.{m}_{bed}^3}\)) flowing from the high concentration at the source (liquid bridges, solids) to low concentration in the gas used to fluidised the bed. The transport through each environment requires a fall in concentration that is proportional to the “current”,ΔCi = MTRw × Ωi, where Ωi is the a called a mass transfer resistance which has the units of seconds. The overall driving force (equivalent to “voltage”) is the sum of all these ΔCi and equals the concentration difference between the source of the water and its final sink, the fluidising gas:
$${C}_S^{\ast }-{C}_g=\sum \Delta {C}_i= MT{R}_{w}\times \sum {\varOmega}_i$$
(10)
The above equation shows the total mass transfer resistance Ωtot = ∑ Ωi to be the sum of the individual resistances, in similarity with Ohm’s law. As the residence time of the gas is short (< 100 ms) it is common in most fluidised bed models to assume that the mass transfer resistance Ωtot and the bed’s temperature and moisture content are constant on the time scale required for the gas to flow from the bottom to the top of the bed. A mass balance over a horizontal slice of the bed with volume dVbed requires the gain of water in the gas flow (φg dCg) to be equal to the mass transferred from the granules to the air (MTRw dVbed):
$${\varphi}_g\ d{C}_g={MTR}_w\ d{V}_{bed}={MTR}_w\frac{1}{f_s{\rho}_s}d{m}_s$$
(11)
Here fsis the volume fraction solids in the fluidised bed (estimated at 40%), and the ρs solids skeletal density (averaged at 1500 kg/m3). Substitution of Eq. 7 in 11 and integration yields (see supplementary material):
$${C}_{g. out}-{C}_{g. in}={f}_{MTR}\ \left(1-{S}_{in}\right)\ {\mathrm{C}}_S^{*}\kern0.75em with\kern0.5em {f}_{MTR}=\left(1-{e}^ {\large -\frac{m_s}{\Omega_{tot}{f}_s{\rho}_s{\varphi}_g}} \right)$$
(12)
Here \({S}_{in}={C}_{g. in}/{C}_S^{\ast }\) is the degree of saturation of the inlet gas which varies between 0 (no water) to 1 for an inlet gas in equilibrium with the water in the granules. It is important to realise that the saturation of the inlet gas (Sin) may change during processing, as \({C}_S^{\ast }\) varies with both bed temperature and water content. We estimated Cg. in such that the outlet air is saturated at the beginning of the constant drying rate period.
The molar drying rate \({\dot{N}}_w\) in the fluid bed dryer now follows from the air flowrate φg and the concentration change calculated from Eq. 12:
$${\dot{N}}_w=\left({C}_{g. out}-{C}_{g. in}\right)\ {\varphi}_g={f}_{MTR}\kern0.5em {\dot{N}}_w^{\infty}\kern0.75em \mathrm{with}\ {\dot{N}}_w^{\infty }=\left(1-{S}_{in}\right)\ {C}_S^{\ast }\ {\varphi}_g\kern0.5em$$
(13)
fMTR is the extent to which mass transfer is limiting: when fMTR = 0 the mass transfer resistances are high and no significant transfer occurs. If on the other hand fMTR = 1 then mass transfer is instantaneous, and the gas phase leaves saturated resulting in the maximum drying rate \({\dot{N}}_w^{\infty }\). The 2rd row in Fig. 11 shows the molar and maximum drying rates. The drying rate is about 80% of the maximum drying rate in the constant rate period which ends at Fw ≈ 0.35, after which the rate reduces in a manner that appears proportional with Fw. Conversely, the maximum drying rate remains stable at Fw < 0.35, as Tbed and \({P}_W^{*}\left({T}_{bed}\right)\) increase balanced by a reduction in the water activity aw as water is removed. The reduction of aw becomes dominate when Fw < 0.1 the driving force and drying rates reduce then sharply.
Step 4 Calculate the Overall Mass Transfer Resistance
The mass transfer resistance Ωtot follows from the ratio of the molar drying rate \({\dot{N}}_w\) measured by NIR, and the maximum drying rate \({\dot{N}}_w^{\infty }\) that follows from the bed temperature and the gas flowrate:
$${f}_{MTR}=\frac{{\dot{N}}_w}{{\dot{N}}_w^{\infty }}=\frac{{\dot{N}}_w}{\left(1-{S}_{in}\right)\ {C}_S^{\ast }\ {\varphi}_g}=1-{e}^{\large -\frac{m_s}{\Omega_{tot}{f}_s{\rho}_s{\varphi}_g}}$$
(14)
The measured temperature and water content data combined with the aggregate isotherm and the vapour pressure of water allows calculation of fMTR. The overall mass transfer resistance then follows by rearranging
$${\Omega}_{tot}=-\frac{m_s}{Ln\left(1-{f}_{MTR}\right)\ {f}_s{\rho}_s{\varphi}_g}$$
(15)
This is shown in row 4 of Fig. 11. After the steady state is reached, Ωtot ≈ 0.03s, but once Fw drops below 0.4, the mass transfer resistance starts to increase, eventually it is an order of magnitude higher. This behaviour is observed in all batches (Fig. 13). At 600 L/min, the curves of the different repeat batches are consistent, even though the 20% relative error in moisture content level does result in significant fluctuations around the mean. The increase in the internal resistance by a factor 30 to 70 is clearly visible in all batches displayed bar Bx 11.
Step 5 Evaluation of the Observed Mass Transfer Resistance
To parametrise, the observed mass transfer resistances we based on a constant resistance external to the granule, and an internal granule resistance that falls exponentially with moisture content:
$${\Omega}_{tot}={\Omega}_{ext}+{\Omega}_{max}\ {e}^{-\frac{1}{F_{W_{crit}}}\left({F}_w-{F}_{w. end}\right)}$$
(16)
Here Fw. end is the final moisture content relative to madsorb. Table III and Fig. 14 shows these parameters for the experiments conducted. As expected, the external resistance varies with airflow reducing from 0.028 s ± 15% at 600 L/min, to 0.013 s±27% for 850 L/min. The maximum resistance (0.64 ±33 % ) and the critical moisture content \({F}_{w_{crit}}\) (0.10 ± 27 % ) appears to be independent of the flowrate. The external resistance will dominate at moisture contents above Fw = 0.5 as only e−(0.5 − 0.1)/0.1 ≈ 2% of the internal resistance remains (~0.014 s).
Table III Average Mass Transfer Resistances for the Placebo Granules The obtained mass transfer parameters are difficult to reconcile with literature, as generally only the drying rate is reported (33,42). It is worth noting that the variation in the estimated parameters related to mass transfer is double the error in the NIR based moisture content (±13%). Even so, the mass transfer analysis using low cost NIR sensors is able to detect the change from externally controlled mass transfer (the so called constant rate period) to mass transfer limited by the internal resistance of the granule. The rate of change of the internal resistance is exponential, which is inconsistent with a shrinking core model in which the volume of the granule that contains water shrinks towards the core of the granule (43). Further work with controlled humidity will be required to see if the analysis is robust.
A detailed phenomenological interpretation of the presented results is beyond the scope of this paper, but it appears that the desorption kinetics dominate mass transfer once the binder liquid droplets have evaporated. It follows that the ingredient selection and adsorption characteristics can have a profound effect on the drying time required.