Testing a Model of Extraction Dynamics

In a recent post, I presented a mathematical model for the dynamics of coffee extraction that is based on a few simple hypotheses. One of these is that the rate of extraction decreases exponentially. The rate at which it decreases can depend on many things: the type of coffee, the roast, the amount of agitation and the brew method would be examples that affect how fast the rate of extraction goes down. For example, in an immersion brew, water becomes gradually more concentrated, and it thus becomes less efficient at dissolving coffee. We should thus expect the rate of extraction to decrease faster in an immersion brew, compared to a percolation, even with the same exact coffee. How fast the rate of extraction goes down is what we technically call a free parameter of the model: the model itself makes no prediction on its value, and can accommodate any number.

One thing that the model crucially lacked until now is being tested against real-world data. As a scientist, I don’t like to leave something floating without testing it, and I want to see how many teeth it loses when it faces the real world. Hence, I didn’t lose too much time and tried to apply it to the data recently gathered by Barista Hustle. As mentioned earlier, an aspect I did not explicitly include in the equations of the model is the effect of fines. The maths behind them is boring, because I make they get extracted to the maximum as soon as they touch water, but I do include their effect to model real-life data. My model makes no prediction about how many fines there are in a given brew, so their proportion is yet another free parameter.

For all those of you new to this, what I’ll attempt here is called fitting a model to experimental data. It’s a game we often play in science: build a model based on some assumptions, which has a few free parameters (think of them like knobs you can adjust to get a different result). Take a set of real-life data, and try to reproduce exactly the same data by playing with your free parameters. If the model is good, you’ll be able to adjust the free parameters such that the model looks a lot like the data. If you have a really large number of free parameters, you will be able to reproduce any kind of data, and your model becomes very poor at making any kind of predictions – in that situation you will even be unable to test whether your starting hypotheses were good or not.

This might all seem a bit abstract, but I think it has the potential to unlock some important understanding about coffee brewing, which I hope will inform us on new ways we can experiment with brew methods and recipes.

Because we will now explicitly work with the Barista Hustle data, I want to remind readers of what their experiment was. They ground some coffee, and sifted it with a Kruve sifter set up with the 250 micron and 500 micron sieves. They thus ended up with three groups of coffee grounds; those that went through the 250 micron sieve (i.e., with diameters smaller than 250 micron along at least one axis); those that went through the 500 micron sieve but not the 250 micron sieve; and those that couldn’t pass even the 500 micron sieve. Depending on how long they sieved, and how much static electricity was present in the grounds, we should expect that some grounds fine enough to pass a given sieve might not always have passed it, but the global result will still be that they end up with three piles of grounds, which average sizes will be (1) smaller than 250 micron, (2) between 250 and 500 micron, and (3) above 500 micron.

They then weighed the exact same amount for each kind of grounds, and placed them in three distinct cupping bowls. They added hot water, and took samples out of the cupping bowls at distinct times. They then measured the concentration of their samples, and worked out the corresponding average extraction yield in the usual way. They got the following result:

So now, the question is – can my model fit this data ? What we have here are 18 data points, but those at t = 0 are really a given (no water = no extraction), so we really have 15 data points to model. When you play the game of model fitting, it is very important that you have less free parameters than the amount of data points. In my case, the model has 9 free parameters, here’s what they are:

• The average characteristic time scale of extraction (τ)
• The characteristic depth that water reaches in coffee particles (λ)
• The maximal extraction yield that this coffee can reach
• The average diameter of particles in the first bowl
• The average diameter of particles in the second bowl
• The average diameter of particles in the third bowl
• The mass fraction of fines in the first bowl
• The mass fraction of fines in the second bowl
• The mass fraction of fines in the third bowl

The first three parameters are required to be the same for all three cupping bowls, because I assumed they use the exact same brew method, and the same agitation in all cases. We also know they used the same coffee and water in all cases. Now, I want to remind you of the hypotheses my model relies on. Any one of these being false will potentially hurt the model’s ability to represent the data.

• The rate of extraction decreases exponentially for each coffee cell.
• The rate of water contact decreases exponentially with the depth of a coffee cell.
• All coffee particles are perfect spheres (the model actually doesn’t change much if you remove that hypothesis).
• Coffee cells are cubic with a side of 20 micron.
• Each bowl contains some fines plus a set of perfectly uniform coffee particles.
• All available chemical compounds get immediately extracted from fines.

We in fact know that some of these must be false (e.g. spherical particles), but it will be very informative to see how well the model fares despite this. Make no mistake: making simplifying assumptions is a very powerful tool in science. It allows you to verify which aspects of an experiment are most important in explaining the outcome, and which aspects have a lesser impact. It is important to bear in mind that the model is a simplified version of reality, but it does not make it useless.

Now, we need to use a recipe to adjust the 9 knobs (the free parameters) of our model in a way that makes it look the most like the data. There are several ways to do this, and this is one aspect of science that I have a decent amount of experience with – a method I really like for this is called a Markov Chain Monte Carlo sampler. The details of it are quite technical, so I won’t go into them. Instead, I will provide you with a simple analogy, that I think is kind of accurate. Imagine the model is a blackbox machine with 9 knobs, and the goal is to adjust each knob individually to reproduce the data the best you can.

A Markov Chain Monte Carlo sampler is a bit like having a hundred monkeys with their own respective blackboxes, randomly tweaking the knobs and judging if the data fits the model. When the model of one monkey starts resembling the data, it starts getting agitated and yells, which gathers the attention of some other monkeys. They get jealous and notice how the successful monkey has adjusted its knobs, and they try to adjust theirs in a similar way, but they don’t do it perfectly. Some monkeys are harder to distract and they keep exploring very different combinations of knobs, until there is a turnaround point where so many monkeys are yelling that really all monkeys are starting to converge on a similar set-up. This moment is called the end of the burn-in phase in technical terms. Once it is reached, you can let the monkeys keep playing with the knobs for a while, and carefully pay attention to what knob combinations they try. At that point, most of what they try will be very close to the best solution. If you gather enough observations despite all the yelling, you will be able to tell what their average knob setting was, and how much they swung each knob around after the burn-in phase. From mathematical considerations, these two aspects will correspond to your best parameters as well as measurement errors for each parameter value. This method is very powerful at exploring all different combinations of knobs, while paying more attention to the combinations that produce better results.

So, I let my hundred computational “monkeys” try 4000 combinations each – the point where all of them were yelling loudly was reached well before the first 1000 combinations, so I paid attention to the last 3000 combinations to determine what the best parameters were. Here’s what the best combination generated:

In all honesty, I was really surprised at this – I expected the model to do much more poorly. You can see that the first cupping bowl with the coarser particles (lower average extraction yields) does not fit as well as the other ones, so whatever effect makes the model imprecise is more pronounced for coarser particles. I suspect this may be related to the fact that each cupping bowl has a distribution of particle sizes and shapes, rather than a very uniform set of particles. Now, let’s have a look at what the best values were for the free parameters, and their measurement errors:

• Maximum extraction yield: 24.0 ± 0.1 %
• Characteristic extraction time: 5 ± 2 s
• Characteristic depth reached by water: 35 ± 8 micron
• Average diameter of particles in bowl 1: 1400 ± 300 micron
• Average diameter of particles in bowl 2: 420 ± 70 micron
• Average diameter of particles in bowl 3: 140 ± 60 micron
• Mass fraction of fines in bowl 1: 49 ± 4 %
• Mass fraction of fines in bowl 2: 50 ± 10 %
• Mass fraction of fines in bowl 3: 50 ± 30 %

The characteristic depth reached by water is smaller than what Matt Perger estimated (100 micron), but I am using an exponentially decreasing reach of water, while he assumes that water accesses the coffee cells equally well down to 100 micron, and then not at all in deeper layers. Matt’s estimation of a 100 micron depth corresponds to the layer where only 6% of water reaches in my model.

The characteristic extraction time is very short, at 5 seconds. This means that, if you were to leave intact coffee cells in a cupping bowl in direct contact with water (i.e., each coffee particle would not have any coffee cell hidden under a surface), you would extract ~63% of all available compounds in just 5 seconds ! This is illustrates of how important it is to consider the effect of coffee cells being hidden under the surface of a coffee particle, as we need much more than 5 seconds to be satisfied with a cupping bowl.

To me, the most surprising parameters were the mass fractions of fines in each cup. They are huge, and almost constant across particle sizes ! I was tempted to make the assumption that each coarse coffee cell has a thin layer, half a cell thick, with broken coffee cells that act like fines. But what we have here is something entirely different: a whopping 50% of all coffee mass seems to be trapped in fines that extract immediately, even in the cupping bowl with coarse particles ! Here’s a hypothesis that I think could possibly explain this – I did not come up with this, but saw a comment that Scott Rao made somewhere about this: a lot of fines may be sticking to the surfaces of coarser coffee particles, possibly by static electricity. It’s also possible that some fines did not have time to migrate through the sieves during Matt’s experiment, even though they were freely hanging out among the coarser particles.

Now, it would be unfortunate to just stop here. The main driver behind why I started worrying about the dynamics of extraction was the question of flavor profile. Since we do not have sensory data to play with, the best we can do is approximate it with a profile of extraction yields. Now that I have a model in which I know how many particles there are of each size, and how fast each layer extracts, it becomes possible to build a distribution of extraction yields for each coffee cell. An even more interesting thing we can look at is the extraction yield profile of each drop of coffee in the resulting cup. To obtain this, we just have to look at how much mass was extracted from each coffee cell, and what its corresponding extraction yield was. We can do this for each cup separately, and at each moment where Matt sampled the cup. Here’s what you get at the first sample collection (15 seconds):

Don’t be surprised that nothing shows up at 0% extraction yield ! This is a distribution of what actually landed in the cup of coffee. While a lot of coffee cells were extracted at 0% because they were near the core of a coffee particle, the “brew” that is 0%-extracted just did not contribute to the cup of coffee. What you are seeing here is a combination of two things: (1) a profile peaking at ~21% extraction yield with a long tail to the lower extraction yields, which corresponds to the stuff that was extracted by diffusion (you might recognize these shapes from my last blog post); and (2) a large peak at ~24% extraction yield, corresponding to the fines which immediately extracted by erosion.

There is something shocking to me about this distribution: between 70 and 80% of the liquid in the cup comes from coffee cells that were fully extracted ! This lends a lot of credence to Matt Perger’s claim that high extractions do not necessarily taste bad, as well as Scott Rao’s comment that fines play a crucial role even if you sift your coffee. This is however confusing to me for one reason: where does all the bitterness and astringency come from in over-extracted brews, if a cup mostly extracted at ~25% tastes good ?

While I do not yet have an answer that I find satisfactory to this question, here are a few hypotheses that I’d like to throw out here:

• Maybe the bitter and astringent compounds have a much slower extraction speed and account for a very small fraction of the mass.
• Maybe the presence of bitter and astringent compounds is explained by something else than high extraction yields. That something else would need to correlate with extraction yield, because we know that astringent and bitter cups have either a higher average extraction yield, or a more uneven extraction yield which caused a larger fraction of the brew to be highly extracted. This is closer to Matt’s explanation that you get bitterness when you “beat up” your coffee too much. No offense to Matt, but I’d really like to find a more precise and scientific description of this process 🙂
• Maybe something is flat wrong with my model, and the fines are not the entire explanation for the very quick rise in average extraction yield in the first sample at t = 15 seconds. It would then be surprising that the model reproduces the data quite well.

Whatever the answer is, I think we need to do more experiments like this one. Having a much finer time sampling especially at the start in our data collection, and going to much longer times especially for the coarse cupping bowl, would be super useful to get better constraints on what’s going on here. This was extremely illuminating to me, but as usual in science, it brought a few answers and many more questions !

If you’re curious about the extraction yield distributions at later times, here they are:

I’d like to thank Mitch Hale, Can Gencer, Mark Burness and Scott Rao for useful discussions.

The header image is by Alexandre Bonnefoy at Issekinicho Editions, Strasbourg, France.