Energy demand reduction in pharmaceutical cleanrooms through optimization of ventilation

The high Air Change Rates (ACRs) required for cleanrooms makes them energy intensive. This research elaborates on three strategies for energy efficient ventilation in pharmaceutical cleanrooms: Fine-tuning, Demand Controlled Filtration (DCF), and optimizing airflow pattern. To study the possibilities for finetuning and DCF, two case studies were investigated and simulations were performed to assess the potential of both options. Experiments in a demonstration cleanroom were used to examine how an ideal airflow pattern may be achieved in the cleanroom, resulting in a high contaminant removal efficiency. Results showed that DCF could lead to substantial energy savings, up to 93.6% in the specific case study facilities. Besides this, DCF based on occupancy could be implemented with negligible effect on the environmental cleanliness requirements. Fine-tuning, based on particle concentration, required representative measurement of the concentration in the cleanroom. It was more difficult to implement in practice. With respect to contaminant removal efficiency, best results (within the experiments performed, ACR in the range of 16h −1 – 38h −1 ), were obtained when air was supplied without a diffuser above the product area and when the work position was located close to the air extraction grilles. © 2019 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ )


Introduction
Compared to an average commercial building, cleanrooms consume much larger amounts of energy [1][2][3][4] . Pharmaceutical cleanrooms can require up to 25 times more power than non-classified rooms: 1.52 kW/m 2 versus 0.06 kW/m 2 [2] . The heating, ventilation and air conditioning (HVAC) system typically accounts for 50-75% of the cleanroom's total electrical energy use [5,6] . This is due to the high air change rates (ACRs) necessary to achieve the required cleanliness classification, as defined in the European Union Good Manufacturing Practice (EU-GMP) [7] . The EU-GMP cleanliness classification refers to maximum allowable particle concentrations in a cleanroom. In order to supply these high ACRs into the cleanroom, supply side requirements are in place. For the supply, swirl diffusers are used. The intention is to provide good mixing throughout the room, leading to a uniformly low particle concentration in the room [8] . Apart from swirl diffusers, other solutions like unidirectional downflow systems are also employed. design"). Together with the ACR, assuming high filter efficiency and theoretical mixing, the amount of particle generation determines the particle concentration in the cleanroom [14] . The cleanroom's ACR should be based on the specific and actual particle generation in the cleanroom. This can be obtained in existing facilities by so called fine-tuning. In this paper, 'fine-tuning' is defined as adapting the ACR to the measured, in-situ, particle concentration, such that the particle concentration requirement for the specific cleanroom is (just) obtained. A second option controls the ACR based on demand. This is called demand controlled filtration (DCF [4] ). In theory, the ACR can be lowered when no particle generation is present. Controlling of the ACR can be achieved based on a night/weekend reduction, occupancy, or particle concentration in the cleanroom. A reduction of 33% fan speed contributes, for example, to a reduction in power consumption by 66% [15] . Previous studies have shown that DCF can result in annual energy savings of 28% to 72% [13,[16][17][18] . A third option for energy use reduction addresses the ventilation efficiency in the cleanroom. The ventilation efficiency in this context is generally expressed as contaminant removal efficiency ( ε) [19] . A higher contaminant removal efficiency allows for a reduction in the ACR as contaminants (particles) are removed more efficiently. This energy reduction option is of interest when positions and intensities of contaminant sources are well-known [20] . Results of a Computation Fluid Dynamics (CFD) study indicate that ε can vary from 0.68 to 9.4, depending on air supply, exhaust position and diffuser type [21] . This would allow the overall ACR to be reduced by a factor of ∼10.
In this study, energy use reduction of cleanroom ventilation was investigated along the two solution directions discussed above: (A) ACR optimization based on performance requirements and actual use; (B) Ventilation efficiency improvement through appropriate ventilation design of cleanrooms. The main research question was: To what extent, ventilation energy use reduction is possible in current and future cleanrooms. In this research, the focus was on pharmaceutical cleanrooms.

Methods
In order to answer the main research question on energy use reduction in cleanrooms, two clear and separate solution directions were investigated. An investigation into the first solution direction (A), ACR optimization, was performed by (1) monitoring an in-use pharmaceutical cleanrooms and (2) modeling and simulation of such rooms to analyze ACR energy saving options. The other investigation into the second solution direction (B), ventilation efficiency, was performed through (3) experiments in a mockup cleanroom laboratory. For each research activity, the applied methods have been described below.

Monitoring in case studies
Two different pharmaceutical cleanroom facilities, located in the Netherlands, were monitored. One pharmaceutical facility was situated in a hospital (Case study H) and one in a radioactive pharmaceutical facility (Case study R). In both facilities, particle concentration and occupancy were monitored during a period of three weeks. In Case study H, two rooms were investigated. Room I is exclusively used for aseptic preparations in a LAF cabinet. In room II, an ampoules machine is primarily used for filling ampoules that will be ultimately sterilized later on. Table 1 provides a summary of the characteristics of the investigated rooms.
Monitoring was performed with particle counters (PC). Monitoring positions for both rooms of Case study H are presented in Fig. 1 . In Case study R, the particle concentration was measured at one location. This location was based on the area where the highest count of microbes (colony forming units) were measured using settle plates. Microbial count measurements were performed periodically by the company operating the Case study R facility. For Case study H, PC1, and Case study R, measurements were performed with a Lighthouse Remote 2014 (1.0 l/min) for a particle size ≥0.5 μm. In Case study H, PC2 and PC3, a Lighthouse Remote Floor plan with measurement positions for the particle counters (PC) and occupant sensors (OS) in Case study H, room I (left) and room II (right). In the right part of the figure, the circle indicates the defined working area (D) for room II. Particle counter 1 (PC1) was placed at different positions during the measurement campaign, indicated by PC1(#); # = week number. In both floor plans the position of the applied diffusers (see Table 1 ) has also been indicated. 5104 (28.3 l/min) was used. All particle counters have an accuracy of 5% [22] .
In order to investigate how the particle concentration in specific areas in room II of Case study H potentially affected results at PC2, the location of PC1 was changed every week (indicated by PC1(#) in Fig. 1 ; # = week number). PC2 was located near the defined working area (D). For the analysis, the correlation in particle concentration between the different locations of PC1 and PC2 was derived. Correlation was assessed with the Pearson product moment correlation, in Microsoft Excel 2016.
In both facilities, Sensor Development people counters (indicated as OS in Fig. 1 ) were positioned such that they registered people entering or leaving the room. In room II of Case study H, an additional movement sensor was installed that measured movement in the defined working area D ( Fig. 1 ). This was done to assess presence in the working area versus presence in the cleanroom. All measuring devices logged their data every minute.

Simulations ACR energy saving potential for case studies
Simulations were performed in Matlab (version R2015a) [23] . The model development was based on ordinary differential equations (ODEs) for a homogeneous cleanroom ( Eq. (1) ; [14] ): In the model, perfect mixing in the cleanroom is assumed. In reality this cannot be achieved [24] . A validation study was conducted using steady state calculations and analysis of monitored data from Case study H. For validation, the contamination source strength was determined from Eq. (2) [25] : In Eq. (2) , D is the source rate [p/s], t is time [s], C is the particle concentration in the room [p/m 3 ], V is the volume of the room and Q the air volume supply rate [m 3 /s].
A sample result from the validation study has been shown in Fig. 2 . There were several changes in occupancy during the investigated period. The model was able to follow these changes. It was assumed that if the model is able to capture this variation, it will perform well at more stable boundary conditions. The results also indicated that a (near) mixing situation, as applied in the simulations, may be assumed for the case studies investigated.
Two different types of simulations were performed: 1. Lowered ACR (fine-tuning): The ACR was lowered based on the particle concentration measured and the source rate derived from that information. For Case study H (Room II), preliminary simulations indicated that the chosen ACR could be set ten times lower than the current operational assumptions, resulting in an ACR of 2.1 h − 1 . This would still fulfill the GMP requirement for the specific case. As the ACR had to be set that low, it was assumed continuously to preserve a minimum pressure hierarchy. 2. DCF based on occupancy: For case studies H and R, the ACR was controlled based on occupancy. No distinction was made based on the number of people in the cleanroom. The ACR of 21 h − 1 and 20 h − 1 , as determined for Case study H and R respectively, was lowered to an ACR of 6 h − 1 when the cleanroom remains unoccupied for 30 min or longer. In the model, a reaction time of 150 s was considered when moving the air supply actuator to another position [26] . This delay was introduced to represent a realistic (practical) situation.
Both simulations were performed with a contamination source rate that was assessed from the monitored data for both cases and calculated using Eq. (2) . Filter efficiency was assumed as 100% in all cases.
Energy savings were calculated based on a relative reduction in fan power consumption following a relative decrease in fan speed. No actual fan performance was simulated. Instead, the savings were calculated applying the affinity law. This law states that the fan power is proportional to the cube of the shaft speed. This translates into: In Eq. (3) , P is the fan power [W] and Q is the volume flow rate [m 3 /s].

Experiments ventilation efficiency
Experiments were performed to provide more insight in to the second solution direction (B) for energy use reduction in cleanrooms: ventilation efficiency. More specifically, the contaminant removal efficiency was investigated.
For the measurements, a cleanroom (6.  [27] with H13 filter class HEPA filters). The FFUs could be controlled to 50% or 80% capacity, which corresponded to approximately 6.1 10 2 m 3 /h and 9.0 10 2 m 3 /h per FFU. Air was extracted at five wall mounted extraction points (A-E; all positioned near the floor). In the cleanroom, a LAF cabinet was present. During the experiments, this cabinet was not operational. A workbench was situated in one corner.
For the measurements six calibrated light scattering airborne particle counters, complying with ISO 21,501-4 [28] , were used. m). All particle counters logged their data per minute and had an accuracy of 5% [22] . Even though all particle counters used were  officially calibrated, a deviation was noticeable between the counters. To make sure that the different particle counters were comparable, all counters were positioned at one location, for an hour, to measure the particle concentration, with the particle source on (the particle generator). Deviations were detected and a correction was applied to arrive at the same concentrations for all the counters (correction factors in the range of 0.8 till 1.2 were used).
A particle generator (Atomizer Aerosol Generator ATM 226) was applied as source (S) (see Fig. 3 ; height 0.8 m) to continuously disperse particles in the air. The used aerosol liquid was Di-Ethyl-Hexyl-Sebacat (DEHS) [29] . The exact dispersion rate of the gen-erator was unknown. All cases applied the same dispersion rate and relative results have been presented and discussed. An estimation based on the extracted particle concentration indicated that the dispersion rate was in the order of 7.9 10 5 p/s for particle size ≥0.5 μm.
Swirl diffusers were mounted in front of the FFUs for the centerline positioned FFUs (position 4-6) as shown in Fig. 4 . When the ACR was set to 38 h − 1 , three FFUs were active at 80%. When the ACR was set to 16 h − 1 two FFUs were active at 50%. This meant that for the supply air position, there were always two or three FFU positions in use, depending on the ACR. For the air extraction grille position, two different extraction grilles were always used. When an ACR of 16 h − 1 was applied, both air extraction grilles were 50% covered in order to maintain an appropriate overpressure in the cleanroom. The PC2 particle counter position was based on the position of the extraction grille that was open for the specific case. The PC4 particle counter position was based on the position of the source (see Fig. 3 ; PC4(#)).
An overview of the performed cases with the corresponding variables has been provided in Table 2 . The codes that have been used for the cases were composed out of the characteristics of the different variables: From the different cases investigated, Case 1 and 2 were regarded as the reference cases. The setup for these two cases is common in pharmaceutical cleanrooms. In Case 3 and 4, the effect of changing the swirl angle from horizontal to vertical was studied. Case 5 till 9 were cases without swirl diffuser, creating a pronounced air flow underneath the HEPA filter (FFU). In order to find the case that had the lowest particle concentration throughout the room, air supply and extraction position were variables in these cases. For Cases 10 and 11, the effect of extracting the air at both sides of the room was studied. In Case 12 and 13, the contamination source was moved to another position. In both cases it was located closer to the extraction position. Besides that, no workbench or LAF cabinet was located in-between the air extraction grilles and the contamination source (which was the case in Case 5 and 7). At the end, Case 9 could not be investigated due to time constraints and has not been discussed further. All measurements followed a similar protocol. Every session lasted 60 min. During the entire session, the aerosol generator was continuously operational in order to assure a continuous particle production [p/s]. The LAF cabinet was off and had its protection screen shoved upward, so that it functioned just like a work bench. Nobody entered the cleanroom during the tests. Every case was performed at least twice, except for Cases 3 and 4, both of which were performed only once. Results from all measurements have been presented in the results section. During the measurements, every day, the ground and workbench surfaces were cleaned with appropriate cleanroom cleaning materials. The ground was cleaned with Ecolab klerwipes and the surfaces with Medipal Alcohol IPA Wipes. The cleanroom was entered wearing cleanroom boots and further normal clothes.
Data of the last 40 min of each measurement session was used for the analysis. In this time range, the particle concentration was approximately in steady state. The average particle concentration of each counter and the according standard deviation over this time range was determined. The contaminant removal efficiency ( ε) was calculated from Eq. (4) .
In Eq. (4) , C exit is the particle concentration at the extract [p/m 3 ], C s the particle concentration at the supply [p/m 3 ] and C mean the mean particle concentration in the room [p/m 3 ].
In the analysis, results from PC2 were assumed to be representative for C exit . The filter conditions were tested before the start of the experiments. No supply of particles was detected in absence of an internal source. Filter efficiency was 99.95% (H13). Based on this analysis and the expected high particle concentrations in the room, C s was assumed zero in the further analysis of the data. C mean is determined from the average of PC1 and PC3-PC6 (over the 40 min period). Significance (p-value < 0.05) in differences in ε between the cases was determined by Mann Whitney tests (2-sided) [30] applying SPSS [31] version 20 and Matlab (version R2018b; Wilcoxon rank sum test) [23] . Apart from the overall (room) contaminant removal efficiency ( ε overall ), the local contaminant removal efficiency ( ε local ) was determined from the particle concentration for the individual measurement positions in the room (C local ).

Results
With reference to the three applied research methods, as explained in the previous paragraphs, the results have also been presented separately for each method, in a detailed manner, to answer the main research question.

Monitoring in case studies
Figs. 5 and 6 show the measured particle concentration for Case study H Room II, and Case study R, respectively, over a period of three weeks. For the measured low particle concentrations ( < 50 0 0 p/m 3 for particle size ≥0.5 μm), the results may be inaccurate due to the low flow rate of the Remote 2014 (1 l/min).
The time that Rooms I and II of Case study H were occupied during the three weeks of measurement was determined at 1.8% and 3.2% (respectively). When Room II was occupied, the employee remained 55% of the time in the defined working area (D; see Fig. 1 ). The cleanroom of case study R was occupied approximately 22.5% of the time.
In Case study H, Room II, the particle concentration was measured at different locations. The highest correlation was obtained between PC1(3) and PC2 ( R = 0.82). For PC1(2) the correlation with PC2 was lower ( R = 0.19). A continuously lower concentration was measured at PC1(2) when compared to PC2. However, average absolute difference in particle concentration between the PCs was small for the monitoring period (less than 0.5%, when applying non-zero concentration values and the GMP-C requirement as a reference).
In general, both Case study facilities operated far below the desired GMP limits for particle size ≥0.5 μm. When occupied, Room  Fig. 6. Measured particle concentration, Case study R. The line shows the GMP C limit. The x-axis represents the time (in weeks), the y-axis represents the particle concentration (in p/m 3 ) for particles > 0.5 μm. The GMP limit (C; in operation) for this room is 3.52 10 6 p/m 3 . GMP B (in operation) allows a maximum of 3.52 10 5 p/m 3 . Fig. 7. Boxplot of measured particle concentration range as a function of the number of employees in the cleanroom (case study R; * indicates a significant difference when there is one more person in the cleanroom for the specific step (p-value < 0.05; two-sided). I in Case study H had a slightly higher average particle concentration than Room II: 1.37 10 4 p/m 3 versus 1.10 10 4 p/m 3 for particle size ≥0.5 μm. As Room II had less strict GMP requirements (factor 10), it had a higher degree of oversizing than Room I.
There was a visible relation between occupancy in the cleanroom and particle concentration. When there were no employees present in the cleanroom, and therefore no source of contamination, the particle concentration approached zero. As soon as employees entered the rooms of Case study H, the particle concentration increased. The detection of particles in Case study R showed a delay of ∼3 min as this cleanroom had a larger volume compared to the rooms in Case study H. In line with the air change rates applied, a decrease in particle concentration could be noticed in all situations within 3 min after leaving the cleanroom unoccupied.
From the measurement data for Case study R the particle concentration, as a function of presence could be derived ( Fig. 7 ).
The obtained data indicates no specific relation between the particle concentration and the occupancy. An increasing overall  Case study H). The x-axis represents the time (in minutes), the left y-axis the particle concentration (in p/m 3 ) and the right y-axis the occupancy (in number of persons). average particle concentration for the first three persons in case study R was notable. The rise in particle concentration was significant for the first three persons (Mann Whitney tests). There was however no significant increase in particle concentration when the cleanroom was occupied by 3 to 6 persons (two-sided).

Simulations ACR energy saving potential for case studies
Following the measured particle concentrations ( ≥0.5 μm), as shown in Figs. 5 and 6 , the simulation results conclude that both for the cleanroom in Case study R and Room II of Case study H, theoretically, the ACR could be lowered by a factor 10 without increasing the exceedance of the GMP C ≥ 0.5 μm concentration limit. Fine-tuning, in this case simplified, to a fixed ACR to (just) fulfil the GMP-requirement and maintain some pressure hierarchy. This would result in energy savings in the order of 99.9%. Fig. 8 presents a sample result of the simulated particle concentration when DCF is applied. The particle concentration in the cleanroom reduces to (nearly) zero within 30 min after occupants leave the room.
In Case study R, when an occupant enters the room, an increase in the particle concentration was detectable after approximately 3 min. This is due to the time required for the HVAC system to switch the ACR. The desired ACR was reached after 150 s. Other causes are found in the size of the room and the position of the particle counter. This implies that a source assumption is only visible after this time interval. This time interval was also assumed in the simulations. In DCF, the control is not on the particle concentration but on the presence in the room. Therefore, the system is already running before actual particles would have been detected. In Room II of Case study H an increase in particle concentration was noticeable faster, due to the smaller capacity of the room. As the HVAC system delay was also included in these simulations, this may result in a higher particle concentration in the first couple of minutes, as illustrated in Fig. 8 . Table 3 presents the final simulation outcomes with respect to potential energy savings, when applying DCF for the investigated case studies.

Experiments on ventilation efficiency
A 2D visualization of the airflow pattern for all cases investigated in the experiments on ventilation efficiency has been provided in Fig. 9 . Table 4 presents the measured average particle con-  Table 4 Average measured particle concentration ( ≥0.5 μm) and standard deviation (SD) for each individual measurement position (for all investigated cases), related to the experiments on ventilation efficiency. The overall contaminant removal efficiency ( ε overall ) for each case is derived from the measurement outcomes for the particle counters, centrations, the standard deviation (SD) and ε overall for the investigated cases. ε local values have been shown in Table 5 .
In Table 5 the values of a single test have been presented separately when a case was performed twice. For ε overall the accuracy of the outcomes was determined at 12% when averaging all outcomes. This value was mainly attributed to the cases with no swirl diffuser, which generally result in a less mixed situation. Looking at the cases with a swirl diffuser the accuracy was 6%. Locally, the difference between repeated measurements for ε local may find larger values. Again, these differences were found mainly for the cases where no swirl diffuser was applied. The measurement position and local disturbances in that case, due to larger gradients, may have affected the outcomes.

Discussion
In this research three different methods were used to investigate how to optimize the ventilation in pharmaceutical cleanrooms in order to derive the energy savings potential. The following sections discuss the outcomes with respect to the main research  Table 2 . question posed. To this end, the monitoring and simulation results have been combined to address the ACR optimization. The outcomes for the ventilation efficiency are discussed separately.

ACR optimization
The monitoring results for Room II of Case study H (see Fig. 5 ) indicated that the room approached GMP B cleanliness almost all the time, except for 6 min in total during the 3 weeks of monitoring. This, in spite of the fact that GMP C demands were in place. The correlation outcomes assumed that local conditions at the work area were not similar at all investigated positions in the cleanroom. So (theoretical) complete mixing was not obtained. Absolute differences, however, remained small. The simulation outcomes, as presented in Fig. 8 , support that a mixing assumption would still be valid for this room.
Case study R showed outliers that were substantially higher than the average concentration in a time period (see Fig. 6 ). These peak moments generally were 1 min samples and may refer to local disturbances. If we do not consider these outliers, Case study R was operating most of the time (99.2%) on GMP B cleanliness when the cleanroom was in use. Room I of Case study H aligned with the requirements set for the specific room. Actual use time of the cleanroom was however very low.
The results indicated that the ACR applied in the monitored case studies generally was much higher than needed for the requested performance. The simulation results for the GMP C cases investigated showed that a reduction by a factor of about ten would still result in a functional cleanroom. This degree of oversizing was also found in preliminary internal research by the authors in the at-rest state [32] . The results presented here extended that to the in-operation phase for the case studies (GMP C).
It is not recommended to apply these low ACRs for the investigated case studies as it would affect the airflow pattern. With the high flow rates, a mixing situation can be expected. When lowering the flow rate, temperature differences will increase as the air is also used to control the thermal conditions. As a result, buoyancy forces may impact the flow pattern more considerably. This was not further investigated. Additionally, the flow must also assure an overpressure in the room. A minimum ACR of 6 h − 1 is proposed in [2] to address these issues. With this proposal, fine-tuning (ACR in the order of 2 h − 1 ) should not be considered for these cases. Lowering the ACR continuously from 20 h − 1 to 6 h − 1 results in potential energy savings of 97.3%.
As the use of a cleanroom is unpredictable and the particle concentration can rise quickly due to activities, control based on particle concentration is less straightforward. There will always be a delay which may result in an overshoot. Furthermore, the particle size of ≥5.0 μm and microbial contamination were not considered in these measurements. Although the research was focused on particle size ≥0.5 μm, the equipment does provide an indication that the concentration of particle size ≥5.0 μm sometimes reached the limits of the GMP C demands. Due to the low flow rate (1 l/min) of the particle counters, accuracy however was compromised at the low concentrations for this particle size (GMP C: 2.9 10 4 p/m 3 ).
The fact that Room I and Room II in Case study H result in similar particle concentrations was explained based on both rooms being occupied by two persons. Room I is twice as small as Room II and has an ACR of 42 h − 1 compared to 21 h − 1 in Room II. This results in approximately the same amount of flow rate (1276 m 3 /h and 1351 m 3 /h respectively). With approximately the same source and flow rate, the same concentration is reached. This confirms that designing based on prescribed ACR does not immediately classify the room's cleanliness, as also stated by Birks [12] .
The influence of more than three persons on the resulting particle concentration in Case study R was not detected (not significant [two-sided]; Fig. 7 ). This may be due to the specific activities performed in the specific cleanroom. It, however, also signifies the difficulty of identifying the particle generation rate when designing a cleanroom. Oversizing of the ACR in that case is assumed in practice to assuredly be on the 'safe side'. For energy efficiency, this is not advocated. ASHRAE [33] recently issued a report which provides more information on particle generation rates and measurement methods to determine this rate. The authors of the report regard it as an initial step and more work needs to be done. As control based on particle concentration (fine-tuning) may not guarantee minimum performance in any case, use of DCF based on occupancy seems a better alternative. In that case the cleanroom functions at its designed flow rate in case of presence in the cleanroom. The number of people present is not a criterion. It assumes that activities in the cleanroom only take place when people are present and particle sources are directly related to occupancy. No sources need be assumed when the cleanroom is unoccupied. The control then simplifies considerably (occupied/unoccupied: ON/OFF), as do the investment costs [16] . If a production process has activities which are not directly coupled to persons present in the cleanroom, control on particle concentration may still present a feasible alternative.
The calculated DCF energy savings for Case study H ( Table 3 ) were higher than in other studies [13,[16][17][18] . The main reason for this was that the two rooms of Case study H showed a very low use time. This was in contrast with Case study R that had a more regular occupation pattern. The energy savings for Case study R were more representative and therefore more in line with DCF particle counting studies from Faulkner et al. (60%) [13] and Tschudi et al. (72%) [17] .
Although simulation results when applying DCF were solely based on particle size ≥0.5 μm, this will not cause any problems for particle concentration size ≥ 5.0 μm and microbial contamination. If a particle concentration of (near) zero is measured for particle size ≥ 0.5 μm, the particle concentration for particle size ≥5.0 μm and microbial contamination, by definition, cannot be larger. The study however does not distinguish the fact that in case of particle generation, due to presence and activity, particle concentrations for particle size ≥5.0 μm and microbial contamination may get closer to limit values than the particle concentration for particle size ≥0.5 μm. This is due to the fact that the particle sources present may produce different distributions of particle size particles. This potential uncertainty provides another challenge for applying fine-tuning.

Ventilation efficiency
Reference Case 1 and 2 showed similar ratios for particle concentration from the different Particle Counter (PC) locations and provided (relatively) homogeneous particle concentrations throughout the cleanroom (see Tables 4 and 5 ). The 2.4 times higher ACR in Case 1, as compared to Case 2, resulted in a significantly ( p = 0.009) cleaner environment. A relatively higher ε local is achieved at PC6 for Case 1 (see Table 5 ) due to the Coanda effect of the swirl diffuser. With the higher flow rate, the applied swirl diffuser was able to maintain a larger clean area in the upper part of the room as compared to Case 2 with a lower flow rate.
A downward directed flow as in Case 3 and 4 resulted in locally lower particle concentrations as compared to the reference case. However, a high particle concentration was measured at PC5 (LAF cabinet), because the contamination is now pushed into the LAF cabinet from the source position (see Fig. 9 ). This is reflected in the low value for ε local at PC5 ( Table 5 ). When PC5 is not considered for the calculation of the overall average concentration, Case 4 provides a significantly ( p = 0.021) cleaner room than Case 2. For Case 3, in comparison to Case 1, this is not significant ( p = 0.386).
In Case 5 and 6, where the air was supplied above the source position with the FFUs (no swirl diffuser), contamination was also directed towards PC5. This resulted in a higher particle concentration at PC5, similar to Case 3 and 4 (see Fig. 9 ). When PC5 was not considered, Case 5 performed better on all measurement positions than Case 6. This is attributed to the difference in position of the extract air grille. For Case 6 the extract grille was positioned further from the source.
In Cases 7 and 8, air supply was located on the opposite side of the cleanroom, as compared to air supply position in Cases 5 and 6 (position 7-9; see Fig. 3 ). For Case 7, this resulted in an improved ε overall , as compared to Case 5 and 6. For Case 8, ε overall improved even further. Compared to Case 6, one could have expected potential short-circuiting to have taken place. This would have lowered ε overall . Instead, the reduced flow rate for this case (ACR = 16h −1 ) may have positively affected the overall flow pattern in the room, bringing it closer to a mixing situation.
The effect of opposite extract positions applying different diffuser types (Case 10 and Case 11) resulted in no difference (nonsignificant) in ε local for Case 10 when compared to Case 2. Both Case 10 and 11 arrive at a similar ε overall when compared to Case 2.
The effect of repositioning of the source was visualized in Case 12 and 13 (PC4 was also replaced in this case). Otherwise Case 12 could be compared to Case 1. When compared, the results indicated that at the workbench (PC1) ε local improved considerably as contamination now was removed more close to the extraction. PC4 and PC3 were located closer to the source in this case and therefore present reduced values.
In Case 13 the supplied air impinged on the desk and was then, for a large part, directed towards the air extraction grilles (see Fig. 9 ). Because the source was located in between the workbench and the air extraction grilles, contamination was removed efficiently. As a result, the highest overall contaminant removal efficiency of all cases ( ε overall = 1.83) was obtained for case 13. In Case 3, 4, 5 and 6 contamination was directed towards PC5 inside the LAF cabinet. This may not be an issue of concern when the LAF cabinet is turned on, as it creates a barrier between the inside and outside of the LAF cabinet. However, when the LAF cabinet was considered to be a normal workbench, the air distribution was not optimal. Since the contamination removal efficiency was only based on the average of five PCs (including PC5), the approximation for the overall contamination removal efficiency was less representative for Case 3,4,5 and 6.
The swirl diffuser setup resulted in a more homogeneous particle concentration, throughout the room. This result was also obtained by Lenegan [8] , using the air change efficiency as a performance indicator. The fact that the 2.4 higher ACR in Case 1 as compared to Case 2 also led to a 2.4 lower particle concentration was consistent with the theory [14] . The derived values for ε overall were at a lower range than the CFD study of Villafruela [21] . This may be partly caused by the ε local values at PC5 for some cases; this affects ε overall . Taking into account the average measurement accuracy for the cases with swirl diffuser, the assumption that turbulent ventilated cleanrooms have an ε overall of 0.7, according to some design guidelines [34] , seems conservative for the investigated cleanroom. Assuming Case 1, ε overall could be set at 0.9 for the investigated cleanroom. In that case, the ACR could be reduced by 22%, resulting in additional energy savings.
Use of localized ventilation, aligned with the activities performed in the cleanroom, was represented in cases with the application of FFUs or swirl diffuser, with vertical downward throw. This can result in locally high contaminant removal efficiencies and therefore efficient ventilation. The measurement results for these cases however do also show larger variation in measurement outcomes with respect to ε local . This implies that the flow pattern is more critical for small disturbances and therefore less stable. A similar remark can be made with respect to the source position. In a cleanroom environment with high requirements on the air quality, such a condition may not be acceptable. In that case, a mixing situation is to be preferred. ISO norms [35] currently also require that all measuring points need to meet the required particle concentration level. This implies that there yet is no benefit to be gained from air distribution configurations with ε local > 1, when at other positions ε local < 1.

Energy savings
The results from the research into ACR optimization and ventilation efficiency improvement showed that, for the investigated cases, large energy savings are possible (in the order of 70-90%). From the results, ACR optimization applying DCF, assuming occupation as control parameter, currently seems the most interesting solution for application in practice. In case cleanrooms have a low use rate, savings can be considerable. Fine-tuning could result in even higher savings but is prone for air quality issues as source control (particle production) is more difficult to obtain and particle concentration measurement at a specific location may not be representative. Furthermore, system response may be too slow. Improvement in ventilation efficiency is possible by localized ventilation, but at room level, sensitivity to disturbances is high resulting in lower air quality. For the investigated case, a mixing air distribution pattern is preferred. The results do however show that current design guidelines with respect to assumed overall contaminant removal efficiency are conservative. For the investigated case, an ACR reduction by 1/5th was possible.

Study limitations
All results shown are for particles ≥0.5 μm. The outcomes may not account, by definition, for particles ≥5 μm due to difference in behavior and deposition.
For the ACR optimization study, the pressure hierarchy was not taken into account. In reality, it is possible that due to a lower ACR a smaller pressure difference may occur and contamination can enter the cleanroom. This potential effect was not investigated further for Case study H and R. For operating theatres, Traversari et al. [36] investigated the effect of switching off the ventilation system during prolonged inactivity. They concluded that the required degree of protection (air quality) was reestablished within 25 min. They however actively contaminated the room in the off-phase. In practice, this will not be the intention. Sun et al. [37] discuss the original assumptions for pressure hierarchy and propose new recommendations. They show that contamination from the corridor (non-clean area) to the clean area is kept at < 0.1% even at near zero (but positive) pressure differences. Follow-up studies may reveal the need to improve/change the leakage area to maintain a required pressure hierarchy to avoid contamination.
The application of swirl diffusers in the experiments resulted in pressure loss as compared to the case with FFUs. As a result the flow rate in case of application of swirl diffusers reduced by approximately 25%. As the particle source rate was fixed for all cases this resulted in differences in particle concentration between the cases (see Table 4 ). The analysis of the contaminant removal efficiency is not influenced by this difference.
The outcomes from the experiments on ventilation efficiency confirmed that the use of swirl diffusers supports the development of a mixing type of ventilation. This supports the mixing assumption that was required for the simulation part. Whether actually full mixing was arrived at in the case studies, was not further investigated.
Due to a limitation in the number of particle counters available, in all cases, data from PC2 has been applied in the experiments for the ventilation efficiency, as being representative for the particle concentration at the extraction point. The assumption is that for cases with the position of the extraction grille on the same side of the room, measurement at one position is representative. This is certainly the case when combined with a mixing situation, as generally obtained with the application of swirl diffusers. Only for Case 10 and 11 representativeness may be argued.
The limitation in number of particle counters was also reflected in the difficulty to arrive at a mean concentration for the room. The focus of the measurements was on the lower region of the room, as here, activities take place and performance requirements are set. In case of mixing, the effect of the position of the particle counters may have affected the mean concentration less. In case of use of FFUs, with larger gradients in the room, the disproportionate distribution of the sensors in the vertical direction may have affected mean values. As the focus is on the lower region, this effect however was regarded as less critical.
Contamination removal efficiency could also have been studied using Computational Fluid Dynamics (CFD). However, this simulation technique requires measurement data for validation. Furthermore, the full complexity of the investigated room, including the supply details, is a challenging endeavor [38] . Nevertheless, the presented results may be used to perform such an analysis. The advantage is that, using CFD, information can be obtained for any point in the room, as opposed to the limited number of field measurement positions. A follow-up study may include the use of CFD to take advantage of that additional information.
Some additional tests were performed with the LAF cabinet on. Because the LAF cabinet included a HEPA filter, it also acted as a particle sink. In those cases, the overall particle concentration throughout the room was significantly reduced, adding to the air quality of the cleanroom environment. An active LAF cabinet therefore may support a cleaner environment. This may be overviewed as another potential energy savings option in the design phase.
Finally, it should be noted that limiting the source's particle emission plays a more important role for environmental cleanliness than the ACR. ISO classes have a factor ten interval with respect to the required particle concentration. Adding two times as much air leads to a two times lower particle concentration when the sources particle emission is the same. Therefore, if possible, it is of primary importance to limit the emission of particles from internal sources in order to achieve a lower particle concentration in the room.

Conclusion
The research shows that for particles ≥0.5 μm, fine-tuning, DCF based on occupancy, and an improved airflow pattern in the cleanroom can contribute to a more energy efficient operating cleanroom.
For the investigated case studies, application of DCF based on occupancy was most promising and has the potential of relatively easy implementation in existing cleanrooms. Reduction of the ACR after 30 min of no occupation is proposed. A minimum ACR is still required to maintain a minimum pressure hierarchy. Applying DCF can result in energy savings for fan use up to 70% and more. These results however are highly related to the actual use of the cleanroom. The savings do not include potential savings for conditioning of the outdoor air. Improvement of ventilation efficiency in a cleanroom is possible by dedicated positioning of supply, source and extraction. Providing the supply air close to the required clean area, the product area, and extracting the air close to the employee's working area can result in a high ventilation efficiency. However, current guidelines do not support such an approach. Requirements are set for the whole room. Therefore, a (turbulent) mixing ventilation flow pattern is still required in case of rooms without full unidirectional downflow. Use of swirl diffusers, with horizontal throw, is a suitable solution for that. The results obtained do imply that current design rules for the contaminant removal efficiency, in case of mixing, are conservative. As a result, for the investigated case, a reduction of 20% in energy use appears possible.
In the design process of cleanrooms, estimation of the (minimum) required ACR, to arrive at an intended recovery time, is significantly hampered by large uncertainties in the quantification of the particle source. This source is determined, amongst others, by the working methods in the cleanroom, the cleaning procedure and the equipment. Appropriate system sizing and potential energy savings would be helped if more information were available on such sources. For now, monitoring of the in-use cleanroom may provide the required information on the particle generation for that specific case. Based on monitoring outcomes, the ACR can be optimized to improve energy efficiency.

Declaration of Competing Interest
None rials to perform the experiments. This paper is based on the thesis from Molenaar [39] .

Supplementary material
Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.enbuild.2019.109346 .