Authored by: James N. Cawse, Ph.D.;Cawse and Effect LLC
Have you been thorough in listing all the factors that might be important?
Factors (and responses) can come from a wide range of mathematical types. This will affect the design, the types of interactions, and the precision of measurement. They range from truly quantitative (real numbers) data through ordinal (good, better best) to binary and finally to non-ordered qualitative data.
There are three major experimental types of control factors:
- quantitative (temperature, flow, concentration,…)
- qualitative (type of material, technique, machine,…)
- formulation (concentration, adding up to a constant such as 100%).
- Normal Level and Range - The known ranges can be used as a starting point to determine the experimental settings. If there are no known ranges, trial runs are appropriate.
- Measurement Precision and Setting Error - DOE calculations are often based on the assumption that there is no error in the values of the control factors. All the error is assumed to be in the responses. Robots are not infallible and should be checked. If the control variables are being set at just two levels, the difference between the two levels should be much greater than the error in setting those levels!
- Proposed Settings - The levels of the control factors should be set far enough apart that it is likely they will have a visible effect - but not so far that the system will be out of reasonable or safe operating range.
A HT program can generally analyze for only one or two responses at a rate that matches the throughput of sample preparation and processing. Definition of the analysis is often the crucial step in setting up a new HT program. Often the analysis cannot measure the customer-critical property directly; instead an easily measurable property that correlates well with the desired property must be chosen. Sometimes a binary response is sufficient. Key Responses information includes:
- Normal operating level and range - The operating results at current factor settings can serve as a reference which allow estimation of the practical magnitude of the effects observed in the experiment. If none are available, a set of trial experiments is definitely in order before embarking on a full HT experiment.
- Speed – The Symyx mantra for HT experimentation is “Analyze in a day what you make in a day”. This sets a stringent requirement for analytical systems. This requirement must be extended to all of the components of the workflow; otherwise the system will choke on its slowest component.
- Precision and Accuracy - Many experimenters do not know the state of control, precision, or bias of their measurement systems. Standard methods are available for bringing statistical control to measurement systems. They should be used.
- Type I and Type II errors. Thought should be given to the error structure of the analysis. In the initial stages of experimentation, analysis which is more forgiving of Type II errors (false negatives) may be a good idea because you do not want to allow potential “hits” to escape. This can be tightened up as the experiment converges on promising regions.
- Relationship of Response to Objective - The correlation of the measured property with the critical property should be known – along with its limits.
What else is going on in the system or its environment? These are the influences which could jeopardize the success of the experiment if they are not understood and accounted for. Key information includes:
- Type - Some factors are controllable but their effects are not of interest in this experiment. They may already have been studied and their best settings are already known. These can be held constant. Other factors cannot be controlled (at least in this experiment). These noise factors can sometimes be measured; in this case they can be treated as covariates. If they can be selected (as in lots of raw material, machine, or operator) they can be treated as blocking factors. If a noise factor can neither be selected nor measured, it becomes part of the causes of experimental error.
- Measurement Precision - It is worthwhile to discuss how well controlled constant factors can be, and how well noise factors can be measured. Any error in the controls or measurements will show up in the overall experimental error.
- Strategy for dealing with factor - The standard tools for dealing with "other" factors are: hold them constant, randomize the runs, and block nuisance factors.
|Click [+] for other articles on||The Market Place for Lab Automation & Screening||Automation Software|