Regulatory Open Forum

In device manufacturing the basic requirement is that your production processes produce only conforming material. If the process were to produce a non-conforming product, the subsequent verification step would detect it and send it to the non-conforming product control process.

However, there are (at least) two cases where this is a problem – the verification step is destructive or the verification step uses a sampling plan.

In those cases, you need a different method, process validation. In this method you understand the process input that affect the process output you need to control. Then, set limits on each process input and monitor them during the production. If the process inputs stay inside the limits, the process output will be only conforming product.

Setting up the initial process validation requires statistical analysis, typically design of experiments methodology. I'm partial to two-level fractional factorial designs. Because there are many input parameters, the input is multi-dimensional. The fractional factorial results identify the worst case corners of the hypercube. After determining them, typically conduct three runs as confirmation. Usually, the worst case conditions are when all the parameters are set high together and when all the parameters are set low together. The third run is the nominal point. The question is the number of replicates in each run. You need to be careful here, because many people use some attribute formula when the parameter is a variable. As a result, you don't get good Cpk information and you run too many items.

In the end you want a production process with a Cpk of 1.33 or better. This provides the high degree of assurance from 820.75(a).

In device manufacturing the basic requirement is that your production processes produce only conforming material. If the process were to produce a non-conforming product, the subsequent verification step would detect it and send it to the non-conforming product control process.

However, there are (at least) two cases where this is a problem – the verification step is destructive or the verification step uses a sampling plan.

In those cases, you need a different method, process validation. In this method you understand the process input that affect the process output you need to control. Then, set limits on each process input and monitor them during the production. If the process inputs stay inside the limits, the process output will be only conforming product.

Setting up the initial process validation requires statistical analysis, typically design of experiments methodology. I'm partial to two-level fractional factorial designs. Because there are many input parameters, the input is multi-dimensional. The fractional factorial results identify the worst case corners of the hypercube. After determining them, typically conduct three runs as confirmation. Usually, the worst case conditions are when all the parameters are set high together and when all the parameters are set low together. The third run is the nominal point. The question is the number of replicates in each run. You need to be careful here, because many people use some attribute formula when the parameter is a variable. As a result, you don't get good Cpk information and you run too many items.

In the end you want a production process with a Cpk of 1.33 or better. This provides the high degree of assurance from 820.75(a).

In device manufacturing the basic requirement is that your production processes produce only conforming material. If the process were to produce a non-conforming product, the subsequent verification step would detect it and send it to the non-conforming product control process.

However, there are (at least) two cases where this is a problem – the verification step is destructive or the verification step uses a sampling plan.

In those cases, you need a different method, process validation. In this method you understand the process input that affect the process output you need to control. Then, set limits on each process input and monitor them during the production. If the process inputs stay inside the limits, the process output will be only conforming product.

Setting up the initial process validation requires statistical analysis, typically design of experiments methodology. I'm partial to two-level fractional factorial designs. Because there are many input parameters, the input is multi-dimensional. The fractional factorial results identify the worst case corners of the hypercube. After determining them, typically conduct three runs as confirmation. Usually, the worst case conditions are when all the parameters are set high together and when all the parameters are set low together. The third run is the nominal point. The question is the number of replicates in each run. You need to be careful here, because many people use some attribute formula when the parameter is a variable. As a result, you don't get good Cpk information and you run too many items.

In the end you want a production process with a Cpk of 1.33 or better. This provides the high degree of assurance from 820.75(a).

In device manufacturing the basic requirement is that your production processes produce only conforming material. If the process were to produce a non-conforming product, the subsequent verification step would detect it and send it to the non-conforming product control process.

However, there are (at least) two cases where this is a problem – the verification step is destructive or the verification step uses a sampling plan.

In those cases, you need a different method, process validation. In this method you understand the process input that affect the process output you need to control. Then, set limits on each process input and monitor them during the production. If the process inputs stay inside the limits, the process output will be only conforming product.

Setting up the initial process validation requires statistical analysis, typically design of experiments methodology. I'm partial to two-level fractional factorial designs. Because there are many input parameters, the input is multi-dimensional. The fractional factorial results identify the worst case corners of the hypercube. After determining them, typically conduct three runs as confirmation. Usually, the worst case conditions are when all the parameters are set high together and when all the parameters are set low together. The third run is the nominal point. The question is the number of replicates in each run. You need to be careful here, because many people use some attribute formula when the parameter is a variable. As a result, you don't get good Cpk information and you run too many items.

In the end you want a production process with a Cpk of 1.33 or better. This provides the high degree of assurance from 820.75(a).