Coffee Data Science
Salami Coffee Grinding
Slicing a grind in time
A salami espresso shot is when you use multiple cups while pulling a shot to examine differences in taste and/or extraction. It is a helpful experiment for understanding how a shot changes in time. I decided to apply the same concept to grinding because I noticed the grind distribution for a few beans differed greatly from 20g of coffee.
These results showed a difference in particle distribution from the beginning to end of the grind, and this could potentially explain why distribution methods that function vertically (like WDT) could improve extraction yield versus horizontal distribution techniques (like OCD).
I used an ice cube tray to great effect, and I collected 13 samples throughout the grind. Then I measured grind distribution for each sample. This first set shows the first two samples, the 8th sample (middle of the grind), and the 13th sample (the last sample).
I found the first few samples are coarser than the rest. Here are the first four samples:
Then the grind starts to stabilize so that by the end, there isn’t much variation between samples in time. Here are the last four samples:
We can look at this data in another way using a cumulative bar plot:
Then if we focus on the cumulative distribution at a few points, the coarseness of the first few samples is clearer.
I wonder if it is due to the grinder just loading up with grinds before it hits a steady state.
Applying Pattern Recognition
Another way to look at this data is by using pattern recognition. I wanted to see if the particle shapes being produced during each sample greatly differed, so I classified each particle using Linear Binary Patterns (LBP), and then I did a K-means Cluster for the particles compared to one another.
With these, I could see based on which clusters each particle belonged to, how similar different samples were:
This shows the first sample is really an outlier even from samples 2 and 3. Samples 4, 5, and 6 seem to cluster well as do samples 10, 11, and 12.
We can split this up by fine (<400um) and coarse (>400um) particles. The <400um similarity matrix is very similar to all the particles because 70% of particles are on the fine size.
What is interesting in the >400um matrix is that there is still such a difference between sample 1 and the rest of the samples. The last sample is similar in that aspect.
We can break these down further in a couple of particle sizes:
300um seems to show a stabilization after sample 3. These rest don’t tell a story much different than the previous.
A Few More Data Points
In some of my other experiments, I have been exploring adding humidity to roasted coffee bean storage, and for six days, I measured the grind distribution from 1 gram and then 20 grams, grinding fresh beans each day. So I was able to compare the first bit and the rest of the grinds.
There was an overall shift towards finer grinds going from the first gram to the 20 gram samples. I plotted it here in another way:
These samples help reinforce that the salami grind samples were not just a one-off, but they were part of how a grinder functions.
This experiment was neat and informative for me. I didn’t think grinding was homogenous, but this was definitely not what I expected.
I wonder if grinding directly into a basket causes issues with grounds distribution being too coarse on the bottom. I suspect this is why a method like WDT or some kind of distribution can help a shot because the actual grind distributions are different in the different layers.
Let’s think about this information in the context of staccato shots. The most basic staccato shots have a finer layer on the bottom and a coarser layer on top. If you grind directly into the filter basket, the grounds towards the bottom would be coarser compared to the top like an inverted staccato shot.
I typically grind into a cup and mix it a bit before scoping it into the filter basket. Usually, I end up with the top of the grounds on the bottom of the filter.
These measurements could also explain why the experiments by Barista Hustle failed to show a difference in weight. They didn’t measure particle distribution, but the real difference in grind distribution is a vertical function.
These results speak more strongly for the need to distribute grounds before tamping whether that is in the grinder or in the filter basket. I don’t know how much coffee clumping plays into the need for distributing grounds after the shot (because I live in a dry climate and rarely have a clumping issue), but certainly grind distribution points to the need for distribution.
First Sample Explanation
Todd Davis in the Home Espresso Aficionados Facebook Group pointed out that the first sample could be an outlier because it is residual coffee grounds in the grinder. I did not clean out the grinder before performing this test, not on purpose but out of forgetfulness. So the first sample could have been a different roast or a different grind setting. Whatever the case, that is the most reasonable and simple explanation, so thank you Todd!
If you like, follow me on Twitter and YouTube where I post videos of espresso shots on different machines and espresso related stuff. You can also find me on LinkedIn. You can also follow me on Medium and Subscribe.
Further readings of mine:
Collection of Espresso Articles
A Collection of Work and School Stories
Measuring Coffee Grind Particle Distribution using Image Processing
A Summary of the Staccato Lifestyle
Measuring Coffee Grind Distribution
Espresso Baskets and Related Topics