Coffee Data Science

Modeling Coffee Grinders

A first stab at objective measures for coffee grinders

Robert McKeon Aloe
Towards Data Science

--

When I first bought a coffee grinder, I didn’t consider if it was good or bad. It was cheap and worked, sort of. It was a Hario hand grinder, and it wasn’t really meant for espresso. After a few other upgrades over the years, I was finally ready to put out some larger bills for a really good electric grinder. The problem I had was not being able to distinguish good grinder from a bad grinder.

My current grinder. All images by author

Most grinder reviews focused on retention or aspects of use related to commercial use. Over the past few years, grinders have been quantified by coffee grind distributions, but those did not give a clear indicator of better quality for grinders over $500. Usually, those tests are with single beans, and the coffee is dialed in.

Who is to say I wouldn’t dial in two grinders differently? What about brew method, water temperature, water flow, dose, or pressure? The number of variables to quantitatively rank grinders is immense.

I have previously scrapped all the espresso related data easily available on the internet, but when it came to grinders, I waited. Recently, I took another look to see if a simple model could predict how two grinders will extract. My model looks only at surface area and an arbitrary to understand how well a grind distribution extracts the good bits and not the bad bits.

Model

When looking at coffee grounds, the available surface area determines how quickly coffee can be extracted. A good way to understand surface area with respect to volume is a ratio (Surface Area (SA)/Volume(V)), and if we assume coffee particles are spheres, then that ratio is 3/radius. For other shapes, the SA/V is usually some constant divided by the radius.

We know that the maximum extraction for coffee is around 30%, so we’re going to say that once you hit that extraction, you won’t be able to extract anymore. Additionally, let’s assume only the first 22% of extraction has most of the good flavors. The last 8% will be assumed to have the more astringent flavors that are less desirable.

Now, all we need is an arbitrary variable to encapsulate flow, temperature, pressure, and anything else influencing extraction. We can assume these area constant, even though flow increases as the puck breaks down. But for the sake of testing, we can sweep across a few numbers for this arbitrary value to see how that impacts extraction. Then we can look at a few different distributions, real and simulated, to see how well this model describes our experience.

Obviously, performing this analysis while also having some real world shot extraction data would be amazing, but baby steps.

Sample

Let’s start with a sample distribution from an EK43 grinder. The percentage of particles doesn’t add up to 100%, but for these tests, that doesn’t matter. All the bins are all relative to one-another. A magnitude shift results in the same results.

First, let’s look at extraction vs modeled rate, and we can also look at over-extraction or bad extraction. As the total extraction goes up, the over-extraction will eventually start going up. Keep in mind that finer coffee grinds will extract fast and thus be the first to over extract.

We can use this Bad EY as a comparison to the total, and we can plot essentially an ROC curve. The best grinder would have a higher Total EY or Good EY for a lower Bad EY. I added theoretical good and bad performance lines to give an idea of how performance would need to change according to the model.

This is a good starting point to a few tests where we can shift the distribution, widen or narrow the distribution, heavily restrict the distribution, or add spikes.

Particle Distribution Shift

We can easily shift the distributions to the right or left to see their effects. The finest distribution will give the highest extraction across rates, but it will also give the highest bad extraction. The coarsest actually has the highest ability to extract without getting the bad parts of the extraction. This is interesting because it is similar to the notion that a coarser ground coffee at a high flow rate can achieve high extraction shots as well.

Narrower/Wider

What happens if the distribution gets more narrow or wider? The effect on the Extraction is hard to see, but the Bad Extraction decreases for a more narrow distribution. This is particularly seen in the Good EY vs Bad.

This idea goes along with what higher end grinders try to do, control the distribution to be in a tighter band.

Restricted Distribution

We can take this concept of narrowing the distribution to the more extreme by keeping it to three bins. This shows little distinction between the 4 variants, but they are much improved over the original.

Spikes

What about spikes? I’ve seen some distributions that have spikes, and spikes on the coarser side don’t hurt as much as the really fine particles.

This test doesn’t have any estimation of taste, but we could loosely assume that a higher EY without bad EY would result in better taste. This also doesn’t take into account the espresso machine used or any other method for that matter.

I’m not sure if this is the best way to view a grind distribution or too simple of a model. However, I think it is a step in the right direction to help answer the question, how do grinders rank against each other?

The next step is to apply this model to available grinder distribution curves to compare and contrast.

--

--

No responses yet