491 Figure 10—SSDEVOP identification of optimal grind time and mill speed in the data of Chauhan (2014). The path of these simplexes was overlaid on the response surface that was initially defined in Chauhan (2014), shown in Figure. It can be seen that the rotation of the simplex indeed took place in the minimum P80 region, at approximately 90 µm. Figure 11 —Overlay of SSDEVOP simplex path and grind P80 contours in the data of Chauhan (2014). In JKTech’s view, SSDEVOP is less preferable to CCRD experimentation. Due to the deliberately incremental changes in variable set-points, SSDEVOP would more often than not, be more time consuming and require more runs. For example, approximately 30 runs were required to identify the P80 minimum in Figure, while the original CCRD required only 13 runs. A variant of SSDEVOP which allows the size of each simplex to expand or contract can, in certain circumstances, allow them to more readily respond to local contour environments (Nelder and Mead, 1965), though this would possibly conflict with the philosophy of targeting deliberately modest set-point changes, to minimise any risk of process upsets. Furthermore, the data generated from SSDEVOP is not as amenable to modelling and delineation of response surfaces compared to CCRDs. In real experimentation, where there is inherent ‘noise’ in the collected experimental data (and with this, uncertainty in the location of the process contours), SSDEVOP simplexes may not stabilise exactly in regions of optimised performance (Paredes and Ágreda, 2020). Figure shows the same surface response as previously shown in Figure, but with experimental ‘noise’ incorporated into the data (defined as SD = 3.5 µm).
RkJQdWJsaXNoZXIy MTM0Mzk2