I just want to add some support for my belief that adding delay time seems to improve the thickness uniformity. I'm not going to into the details about the statistics, but for those who are comfortable with statistics, I'm comparing the means of the various trials in a two tail t test with the null assumption that the trial means come from the same distribution. I've also assume unequal variances. The short description is that I'm trying to be conservative in the testing.
For the trials in the two graphs of 9 second delay and 13.2 second delay, with an alpha of 0.01, the probability that means are the same is far less than 1% (actually about 5 orders of magnitude less). In other words the thickness mean differences are likely real.
I've also compared 13.2 second delay with 15.2 second delay. At an alpha of 0.01 it suggests that the thicknesses are the same, but at an alpha of.05 it suggests that the means are indeed different. Another way to say this is that the evidence that adding two more seconds of delay improves the thickness is weak.
I've probably reached a point of diminishing returns but am running some trials at 17.2 seconds to confirm it.
I have a few more observations on the delay effect on thickness measured. When I extended the delay from 13.2 seconds to 15.2 seconds the respective means and variances are: 5.2925/ 0.006385 and 5.235/ 0.008174. The t test (two tail, unequal variances) p value is 0.024 which is moderate support to think the means are from different distributions.
I also increased the delay to 17.2 seconds which produced mean/ variance of 5.22625/ 0.00592. Comparing this run to the 13.2 second run gave a much better p value of 0.00532 which is by many standards a strong suggestion these two data sets are different. By themselves the differences between the 15.2 and 17.2 second delay data are too small to have any confidence they are different. This data may be suggesting that the optimal delay time for this resin and likely the part geometry itself is greater than 15.2 seconds.
I should point out that these differences (10's of microns) are getting rather small to pull out of the data with a micrometer. There are also only 24 data points generated in the three replicate trials. Drawing conclusions under these conditions is getting sketchy, but I will also follow up with another trial using 20.2 second delays. I do suspect that I beyond a point of diminishing returns but I'm this far in so why not?
02-26-2021, 07:41 PM (This post was last modified: 03-04-2021, 07:25 AM by DDLLC.)
Ran another replicates with a delay of 20.2 seconds and notice something else odd. I thought that the delay amount was appended to the time going up and down, but that was not the case. Build plate stayed at the bottom of the tank for 3 seconds before moving up and then appended 14.21 seconds after touching the bottom of the tank (before the light turned on.) Total light-off light-on time was circa 19.3 seconds; close enough to the programmed 20.2 seconds. I'm not sure if such a division was present in other delay time runs. It could also be related to the screen image turning off before the actual lamps.
The 20.2s run is likely different than the 15.2 s run at an alpha of 0.05. The 20.2s run vs 17.2s run fell just short of passing the Tcritical level at alpha 0.05. In plain english, there is a suggestion that the difference in the means of 20.2s run and 15.2 s run is real, but it isn't clear that going up to 20.2 s delay from 17.2s offers an improvement. For those that understand it, the p stat is 0.069 (close but didn't make the cut).
I've attached a chart that shows results of various average thicknesses versus delay. It shows a trend but at the lower thicknesses with the number of data points and the associated variances, it is probably progressively more nebulous if there is a true improvement. I do believe that increasing delay times improves accuracy and precision.
EDIT
There was a mistake on the standard deviation for 20.2 second delay in the chart posted before. It is now corrected and expanded to include 25.2 second delay. Statistically the data from the 25.2 second trial being different from 17.2 seconds trial is orders of magnitude (about 300x) more likely than the difference between 20.2 second delay and 17.2 being likely. In other words, there is a strong possibility that increasing the delay to 25.2 seconds will produce more accurate and possibly more precise printed thicknesses. Note that the 20.2 stats mentioned prior to the edit are a bit different, but I'm not correcting the above as the replaced chart and the extended data provide better information.
For some reason I couldn't delete the original chart but reference the corrected one.
