Figure 3 shows data from a biomanufacturer for clean-in-place (CIP) times from 2002 to 2008. Those data came directly from a manufacturing execution system (MES) to provide an unbiased estimate of facility performance.
(Bio-G software integrates directly with most common control systems in biotech facilities.) Note the significant variability in this process step, which takes between three and six hours. There is also process drift: The amount of time the activity takes is not constant from year to year.
The data suggest that using a single number — say, the average time — for planning and optimization may not identify the correct bottleneck. My experience in this area has been that running the same model with and without variability produces different bottlenecks. Such an analysis is especially of concern because finding and fixing bottlenecks can be a capital-intensive process when retrofitting a validated facility. “Fixing” the wrong area will not improve run rate no matter how much improvement is made in that particular area.
Another issue that must be overcome in debottlenecking and process optimization is process complexity. Figure 4 shows a map of the places in one facility where people interact with the process. A delay in the availability of a full-time employee (FTE) or piece of shared equipment (e.g., a CIP skid, transfer line, valve array, transfer panel, or even WFI or utility systems) will delay that process. Such “hidden” periods are not normally considered in cycle-time calculations. Complex interactions make it difficult even for highly trained operators to understand which hidden cycle times and waiting periods will slow a facility down.
Our software uses discrete event simulation and real-time data feeds to manage both variability and complexity. Discrete event simulators incorporate variability as well as constraints around how an activity can start. (So just like an automation system, an activity will not start without certain prerequisite conditions met.) This approach produces accurate models of biopharmaceutical facilities that incorporate variability and “hidden” wait times.Bottleneck Identification
The traditional approach to identifying bottlenecks (made popular in the 1960s) focused attention on facility resources that get the most use. The theory was that such resources are “busiest” and therefore those in which improvements would have the most effect. But this approach is not guaranteed to identify the real bottlenecks in a facility. It appears from empirical evidence to be particularly ill suited to biomanufacturing facilities.
Consider as an example a facility with one large buffer preparation tank used to prepare buffer for the first two chromatography steps in downstream purification. Because the tank is used only in the first two high-volume steps of the process, its overall use is low (Figure 6). However, using that tank for two sequential chromatography steps makes the second step wait for it to be cleaned and buffer reprepped. If that process takes longer than the duration of the protein-A chromatography step, it could delay the start of the cation-exchange step. Thus, even though the tank has an overall use of 20%, it delays production and therefore represents a bottleneck.
So a simplistic use-based approach to debottlenecking does not find the true process constraints in a biomanufacturing facility. The current “gold standard” is to use a simulation model to perturb (make controlled changes to) a model of the facility and observe the results according to some metric. One popular approach is to reduce operation times in each process step, one at a time, and observe their impacts on throughput. By simulating a facility in which a particular process step (e.g., CIP) takes zero time, we can examine the effect of having that activity “for free.” By repeating this experiment for every activity, we can understand the chance of reducing a potential project's cycle time. This sensitivity analysis can correctly identify even complex bottlenecks because it makes an actual change to a system model and illustrates the effect of that change.
Figure 7 shows the output from just such a sensitivity analysis. In this case, the metric chosen was throughput, so a higher number indicates better performance. Each point in the graph represents reduction of a particular activity's cycle time to zero hours and the resulting throughput observed over a 100-day campaign. The experiments are sorted with those with the most significant impacts to cycle times on the right and those with the least on the left.
As that analysis illustrates, only a small percentage of the several hundred activities in the manufacturing facility actually improve throughput. Figure 8 shows detail of the four activities that most affected throughput in the Figure 7 model (those on the far right). This approach detects bottlenecks very rapidly and is guaranteed to identify real process constraints in a system because it simulates the actual impact of a change to each constraint in turn.