Coffee Particle Similarity Between the Niche and Ode Grinders using Pattern Recognition | by Robert McKeon Aloe | Jul, 2022

Coffee Data Science

Analyzing shapes using pattern recognition to better understand fundamental particle shapes

In an effort to better understand the differences between conical and flat burrs, I wanted to better understand how they cut coffee beans. My aim is to use pattern recognition on the particles to see how their shapes differ.

Previous efforts were useful in showing the difference between the coarser particles of the Niche and the Rok, so I pulled out some image data for the Niche and Ode, and I went to work. The main caveat is that this is sampled from a single coffee bean at a single grind setting. Both grinders were dialed in to produce similar particle distributions. Results may vary for other beans and settings; more data and study is necessary.

Espresso Machine: Decent Espresso Machine

Coffee Grinder: Niche Zero and Fellow Ode with SPP Burrs

First, let’s plot the distributons. I looked at two settings on the Ode to find the closest to the Niche. The distributions were slightly different, but they have a lot in common for below 300um in particle diameter.

I then used Linear Binary Patterns (LBP) to categorize each particle, and then I used K-means clustering on all the particles to see how they align.

First, we can look at what percentage of the particles are grouped into which clusters. There is a little bit of movement in cluster alignment, but not much.

Then we can compute the average difference in particles of different sizes to other particles using the cluster alignment as a pattern vector. This compares the clustering of all particles in one group to the clustering of all the particles in another group.

It is not unexpected that particles of similar sizes are more similar, but it is a little tricky to tease out the differences. Let’s take a closer look:

The Niche has particle sizes less similar to larger particle sizes than the Ode. When comparing the two, same size particles are pretty similar.

We can take a closer look at same particle size vs dissimilarity score. The lower the score, the more the two are similar. The score stays steady until 400um.

We can also compare the current bin to the next bin of particle sizes to compare the Niche to the Niche and understand how those differences compare to the Ode. Again, there is not much difference between the Niche and Ode until 650um.

Let’s split the particles into two sets for each particle size so that we can better compare a particle bin vs itself for each grinder. Each particle size bin is split into two groups the x-axis uses one group and the y-axis uses the other in the comparison.

When we look at this diagonal, the main difference comares around 550um.

After examining and comparing particles to one another, I did not find sufficient evidence that the two burr sets create fundamentally different particle shapes. Therefore, for this coffee and grind setting, the dominate influence to taste from the burr set is from the particle distribution, not the particle shapes.

Coffee Data Science

Analyzing shapes using pattern recognition to better understand fundamental particle shapes

In an effort to better understand the differences between conical and flat burrs, I wanted to better understand how they cut coffee beans. My aim is to use pattern recognition on the particles to see how their shapes differ.

Previous efforts were useful in showing the difference between the coarser particles of the Niche and the Rok, so I pulled out some image data for the Niche and Ode, and I went to work. The main caveat is that this is sampled from a single coffee bean at a single grind setting. Both grinders were dialed in to produce similar particle distributions. Results may vary for other beans and settings; more data and study is necessary.

Espresso Machine: Decent Espresso Machine

Coffee Grinder: Niche Zero and Fellow Ode with SPP Burrs

First, let’s plot the distributons. I looked at two settings on the Ode to find the closest to the Niche. The distributions were slightly different, but they have a lot in common for below 300um in particle diameter.

I then used Linear Binary Patterns (LBP) to categorize each particle, and then I used K-means clustering on all the particles to see how they align.

First, we can look at what percentage of the particles are grouped into which clusters. There is a little bit of movement in cluster alignment, but not much.

Then we can compute the average difference in particles of different sizes to other particles using the cluster alignment as a pattern vector. This compares the clustering of all particles in one group to the clustering of all the particles in another group.

It is not unexpected that particles of similar sizes are more similar, but it is a little tricky to tease out the differences. Let’s take a closer look:

The Niche has particle sizes less similar to larger particle sizes than the Ode. When comparing the two, same size particles are pretty similar.

We can take a closer look at same particle size vs dissimilarity score. The lower the score, the more the two are similar. The score stays steady until 400um.

We can also compare the current bin to the next bin of particle sizes to compare the Niche to the Niche and understand how those differences compare to the Ode. Again, there is not much difference between the Niche and Ode until 650um.

Let’s split the particles into two sets for each particle size so that we can better compare a particle bin vs itself for each grinder. Each particle size bin is split into two groups the x-axis uses one group and the y-axis uses the other in the comparison.

When we look at this diagonal, the main difference comares around 550um.

After examining and comparing particles to one another, I did not find sufficient evidence that the two burr sets create fundamentally different particle shapes. Therefore, for this coffee and grind setting, the dominate influence to taste from the burr set is from the particle distribution, not the particle shapes.