MINITAB and Design-Expert are two software that features Response Surface Methodology (RSM). It consists of first planning a set of experiments, that is, selecting a Response Surface Design (RSD) first order followed by a second order RSD, secondly, obtaining the experimental data, and, upon collection of data, analysis of the design for modelling and prediction.
For simplicity I would recommend Design Expert. It is so simple to use and understand when it comes to experimental design and data analysis. But constructing 3D surface plot and contour graph for the analysis I always use statistica 12.5. It has the best 3D surface plot, contour graph as well as Pareto chart. Someone can easily explain and understand the combine,effects of variables on their responses when using statistica than using Design Expert. If you are just learning experimental design using Response Surface Methodology, I would advice you to stick to Design Expert. Statistica is little bit complicated than Design Expert. If you want to understand more on how to use design expert, you can check my profile and click on research items you will find a comprehensive handout on how to use Design Expert for experimental design. Thanks
What about the application of response surface methodology on the already obtained experimental data, for statistical modelling.Consider that no DOE has been considered for the experiment and the data is in literature. Is it necessary to construct a Box-Behnken or Central Composite designs? In this case the number of runs usually does not match with the actual experimental observation points. Or is that just enough to apply response surface regression to the data and to get model? How accurate or significant can it be?
Dear Nurlan Amirov, if an experiment had been carried out and result had been obtained without using DOE or RSM design in the initialization stage of the experimental design, RSM cannot be applied to the data obtained from such experimental design. If u force ur data into RSM, the regression model obtained will not be significant. It is always advisable when planning an experiment in which optimization is the Final goal of the experiment to use RSM methodology. The choice of design (BBD or CCD) will depends on the factors and the levels of factors u want to consider in the experimental design. For more clarification on experimental design using RSM, you can send a direct message to me or send an email to [email protected]. Thanks
Babatunde Adewoye is very correct. Before you can use RSM for statistical analysis, you need to use same for experimental design. You cannot force any data not designed using RSM on it
Actually, there are some such model related papers in literature in which the models have been developed on the basis of the data of someone else who did not intend to use DOE or RSM, as far as I understand from the paper, which contains the original experimental result data. In such works the models are satisfactory enough, most of p values are less than 0.05 and the models include several terms also with quadratic ones. But in my case, the p value is a great problem, there is only one or two terms significant. It is quite difficult for me to understand how they get such significant models. I provide the links just for one of such models, but if you would like I can send others. Thanks
Green fuel from coal via Fischer–Tropsch process: scenario of optimal condition of process and modelling - https://doi.org/10.1007/s40789-018-0204-7 and Original data obtained from: Effect of reduction temperature on a spray-dried iron-based catalyst for slurry Fischer–Tropsch synthesis - https://doi.org/10.1016/j.molcata.2006.07.056
In one paper it is clearly stated that "in order to develop RSM models, it is necessary to collect experimental data based on the DOE methods. In the reference work, experimental data were not collected by DOE methods. To compensate for this drawback, ANNs in which original data were trained to produce appropriate input data for RSM were constructed. Then RSM was used for the modeling"
Artificial neural networks are applied in this work to make the data useful for RSM, which have not been collected by DOE methods. But, the thing is that in other works also the data is not DOE obtained, but ANNs are not used in such cases, for example the paper I provided above.
So, what do you think, is ANN the only way to train the data and make it useful for RSM modelling or is there any other way?