From Powder to Precision: Digifabster helps you solve the SLM print time puzzle

A seemingly boring, but very important aspect of real life is statistics. I learned that lesson when the interim function of BD manager was added to my already full plate of New Tech Director at a major glass factory.

A client would come up with a new design, i.e. a new bottle, and would ask what it would cost to produce in a batch of x0.000.000 pieces. This question I would relay to the technical staff, and then would have to wait a month for an answer. During that month the customer would call me at least once a week, getting more and more impatient.

What would be happening in the meantime at the glass factory was a serial orgy of meetings where everybody had his or her say about the new design, would make comments on how to improve it -thereby making it un-new and thus valueless- and do some guesswork at how fast it could be produced, leaving the customer stuck with his marketing and sales plans, not to mention his production plans. And these were really big customers (think Heineken or Carlsberg).

Seeing both sides of the coin I knew I had to find a way out: You can’t keep major customers waiting for a month, and you can’t hurry techies when they have an egg to lay.

So I turned to statistics. Luckily we had by that time “borrowed” a lot of know-how from other factories, amongst which bottle descriptions, planning methods and statistical data. What made the statistical data opaque was the fact that there is a huge number of variations on the theme of “Independent Section (IS) machines, the main difference being the number of cavities (molds) mounted on each machine.

So the first step in making the data usable was taking account of the number of cavities on the machine. Dividing the result per hour for bottle “x” by the number of cavities on the machine in question would result in cycle times per unit of anything from 4 to 12 seconds.

Putting the bottle glass weights in one collumn and their cycle times in another would give a correlation of 0,95, meaning: the heavier the bottle, the longer it would take to produce.

Splitting the bottles, cycle times and weights according to the three main technologies used to produce them, would raise that figure to 0,98. Weight was clearly the key factor.

Now all we had to do was figure out the formula. It turned out that Owens- Illionois had already done that, only they used it not to predict production before the fact, but evaluate it after the fact, same difference. The formula was a very complicated one, describing a hockey-stick curve. In short: a bottle twice as heavy would take somewhat less than double the time to produce. Nowadays I would use our “exponent smaller than 1” DigiFabster formula to describe it, but for the situation in hand a big plot, pinned to the wall of my office, was sufficient.

From that moment on my life became much easier. I would get an email with a bottle drawing with content volume, dimensions and weight, plus the needed quantity per month. The dimensions and weight gave me the technology, the quantity per month the machine, the plot on the wall the cycle time, and thus machine time, thus machine time cost price, the other pricing component being the weight of glass consumed times the cost price of glass. Such a calculation would take me 5 minutes. I would then do two things: Send the drawing to the factory to have the techies brood over it, and tell the customer the price of the proposed contract. After a month I would get the formal confirmation from the tech department, usually no more than 1% off my own prediction.

The customer would by then already have ordered his molds and be printing his flyers and recording his commercials, and most importantly, happily making the down payment on the first production run. Money in the pocket, one month ahead of the usual schedule, just by using statistics.

When after a year the factory engineers found out what I had been doing they were not amused. Well, I wasn’t reporting to them, I was reporting to the owners, and the owners were happy because the customers were happy.

Now why is all this relevant for 3D print quoting?

A few points:

1. People with a stake in one approach (meetings, meetings, meetings) will actively or passively oppose other, less labor-intensive and thus cheaper methods. This is a fact of life, and has to be taken into account with every automation project. There are ways to handle it, the solution includes empathy, empowerment, and showing a (believably) brighter future. For example: None of the engineers in those meetings mentioned above got fired, but the whole innovation process for the factory accelerated after my approach became common practice, simply because hours and hours of engineering time per week were freed up for better purposes. That way everybody became richer faster :-)

2. Don’t be impressed by complication, like multiple configurations, several technologies or lots of different, seemingly unrelated, figures. What we do, 3D printing, is a physical process, ruled by strict laws no amount of marketing blahblah can break. A laser has a maximum installed power, melting an maximum amount of a given powder per unit of time, a drive has a maximum acceleration for a certain weight, also predetermined by the installed power. Given that nobody wants to build expensive machines by installing oversized hardware, those pieces of equipment will be operating in a range from 50 to a 100% of their max, which in combination with other parts working in their own but similar ranges gives quite a stable overall output. Once you know in rough outlines how the thingy works, you can start to predict how fast this or that model will be produced in relation to another model.

3. A common approach to “calculating” machine time for a model is letting the proprietary printer software generate the g-code. That’s probably very exact, but you are doing it in your customer’s time, like my techies at the glass factory who were wasting a month to figure out machine speeds for a new product while the client was cooling his heels in the waiting room. It’s as efficient as looking at a map and trying to work out an ETA based on the speed limits on every section of road. If you’ve been driving through Europe for a few years, like I have, you know that you’ll be doing 90 km an hour on average, so a trip of 900 km takes ten hours. Calculating it in detail would add at least another hour.

4. Another approach, much faster, is to do a number of observations of the relations between model properties and machine time and formalizing them in a spread sheet. That’s absolutely the right way of doing it, and the more observations and properties, the more exact the prediction will be, but you’ll still have to get the properties out of the model and into the spread sheet. Another issue is that after some time and changes of sales operatives, nobody remembers how the original observations were made (or what they were) and the spreadsheet becomes holy writ. This is another situation we at DigiFabster sometimes come up against: people don’t understand how to get their spread sheet formulas into DigiFabster templates because they don’t know how they were arrived at.

The solution in both cases is the same: make new observations, collect the data in a file, and send them to us.

I personally spent the last two weeks taking such a file with 150 models, two technologies — SLS and SLM, on the following machines: EOS’ Formiga P110 and M290, and 4 materials: nylon, titanium, steel and alumina, at two layer thicknesses each, apart. The job was to predict machine time costs for all of those configurations, based on the print time the proprietary printer software had generated from its g-codes.

First of all: my respect for the person or persons who created the file, that must have been incredibly boring. Second: I got a whole new appreciation the job our software does for us, extracting vital model data and imputing them automatically. I had to do that by hand: open model in Meshlab, switch correct filter on, create the right view, copy/paste dimensional data into excel, next…

What I was entering were x, y, z, (bounding box), model volume, and model surface: 5 data points 150 times is 750 data points to enter, by hand. Then I had to split the time notation -which was literally in the format of “12 hours and 25 minutes”- into 12, 25, excel added 60*12+25, which gave me total minutes. Please, if you ever want to send us a file like this, note hours and minutes in separate cells, without any extra text, that gives us something to work with immediately. Now I had to do this manually 150*4*2=1200 times. There is probably a better way, built-in in excel, but I didn’t take the time to figure it out ;-)

At last I had all my ducks in a row and was ready to start experimenting. I was looking for the factor that most influenced print time.

So I put a correlation calculation under the columns which had x, y, z, volume, surface and complexity in them, all in relation to print time.

The results: SLM (and a bit of SLS)

The results: For SLS I got correlations between 0,9855 and 0,9995 for the relation between z-axis (model height) and machine time.

For SLM it was the same range, 0,98 to 0,99, however, not for z, but for volume related to print time.

I will continue the discussion for SLM only, one material and one layer thickness: Ti64 30 micron

Now that I had found the key factors for machine time, I still had to find the relationship itself. By plotting the volume and and machine time data for SLM, Ti64 30 micron I got a typical hockey stick curve. Our program, however, allows only a linear relationship between volume and print time. What to do?

Here I remembered the old proverb: “Time is money”. DigiFabster is not (yet) into ERP, our thing is cost prediction. So I can use money to represent time, and time to represent money, as long as the final cost calculation is correct. Knowing the technology itself and its similarity to SLS I guessed that the next important factor would be growth along the z-axis.

The owner of the machine had told me that the print time cost was 60$ per hour. I did two things: I took the maximum melting speed by dividing all volumes through their respective print time in minutes and using the MAX()-function, thus getting a melting speed of 200mm3/minute, or 12 cm3/hour,and then used this maximum speed to recalculate a print time and print cost: volume/12cm3=calculated print time in hours, @60$/hour.

In parallel I calculated the maximum growth speed, by dividing the lengths of the z-axes by their respective print times in minutes and again selecting the maximum value. The result was 14,3 mm/hour @60$/hour = 60/14,3=4,2 $/mm along the z-axis.

Combining those two: a melting speed of 12cm3/hour @ 60$/hour and a price per millimeter along z of 4,2$/millimeter I got prices which were on average 0,99362% of the original print time, generated by the proprietary printer software, in hours, multiplied by 60$. The standard deviation was 6,6%. A little tweaking of both factors got me to 99,9% and 1,54% respectively.

“Standard deviation” may be a term not completely familiar. It describes the distance of all the found data points from the average, thus predicting a spread of values in a normal distribution. It allows me to state the following: Only 0,27% of the prices generated by this calculation will be more than 4,62% off the price predicted by EOS’ proprietary software.

According to the owner, the printer software print time prediction is +/-2% off the factual print time anyway, and takes roughly 3 minutes per calculation. Our tweaked multiplications, once the factors are input in the DigFabster printer settings, take 3 microseconds, so are roughly 100.000 times faster.

Now let’s have a look at our customer’s current practice:

1. He opens his email.
2. Copies the model in the attachment to his hard disk.
3. While noting at the same time who send the model.
4. He uploads the model to his printer software and waits 3 minutes.
5. He copies the print time calculation and other pricing info, let’s say dimensions, to a spread sheet.
6. The spread sheet gives him a pricing result.
7. He copies the result to an email back to the sender of the model and
8. Gets an approval for the price (or not, in 2 out of 3 cases).
9. Prepares the print.

I see 15 minutes of time wasted right there.

Once this prospect customer switches to DigiFabster, which we hope he will, he saves those same 15 minutes of lethal tedium (and I know what I’m talking about, I just spent two weeks in quoting hell) to do more interesting things, like planning the purchase and installation of another printer.

At the same time he gives his end user superb service: No waiting by the mail box for the quote to come in, but upload, choose, pay, and back to his own customers.

The short version of this discussion, in the form of a step-by-step explanation, has been published in our knowledge base.

Stay tuned :-)

This article was first published in the DigiFabster blog.