I’m a huge fan of IKAWA. I use their roasters every day, and I have a lot of respect for their team. But for this article, I’m putting my fanboy feelings aside and focusing purely on the data. No bias—just an honest look at what this AI-driven first crack detection can (or can’t) do.
This wasn’t some massive research project—it was a small experiment to satisfy my curiosity. Should I rely on AI first crack detection, or is profile replication the safer bet? That’s a question for another article. I also tracked moisture loss, but after the third roast, I realized that some of the lighter beans got spit out during cooling, and I hadn’t added them back in the earlier batches. Since that messed with the accuracy, I decided to leave those numbers out.
Oh, and I didn’t cup these coffees. This experiment was purely about consistency in color—what the IKAWA can do with its AI-driven first crack detection. Taste testing would be the next logical step, but that’s not what this was about. Think of it like measuring the height of two basketball players without checking if either of them can actually dunk.
First crack is like the coffee bean’s version of puberty—it pops, expands, and things start changing fast. Normally, roasters listen for that distinct cracking sound, but the IKAWA Pro100x takes a different approach. It doesn’t have ears—it has sensors.
Real-Time Moisture Monitoring
As the beans heat up, they lose moisture—pretty standard.
The IKAWA’s humidity sensor tracks moisture levels inside the chamber.
AI Analysis of Moisture Peak
Just before first crack, there’s a big jump in moisture release—kind of like an exhale before a sneeze.
The AI detects this peak and marks it as first crack.
Automated Adjustments for Consistency
Once FC is detected, the roaster automatically adjusts the profile, using a pre-set development time.
For this test, I set the development time to 60 seconds after first crack to keep things consistent.
As of the time of this writing this AI first crack detection feature is only available on the IKAWA Pro100x. It can be turned on or off, depending on whether you trust the AI or prefer manual monitoring.
To put this feature to the test, I roasted two coffees four times each:
El Salvador Santa Anna washed coffee: A slow-developing, crowd-pleaser (82-83 points), with a moisture content of 10.5%.
Ethiopia Shanta Golba natural coffee: A Grade 2 coffee with wild bean size variation—which made IKAWA’s job harder. Its moisture content was 10.6%.
The goal? To see if the AI could maintain consistent roast color across multiple batches.
Agtron Ground : 105.1
Agtron Ground : 87.1
Agtron Ground : 93.4
Agtron Ground : 126.6
Agtron Ground : 105.1
Agtron Ground : 87.1
Agtron Ground : 93.4
Agtron Ground : 126.6
Agtron Ground : 121.3
Agtron Ground : 121.4
Agtron Ground : 117.3
Agtron Ground : 124.3
Agtron Ground : 121.3
Agtron Ground : 121.4
Agtron Ground : 117.3
Agtron Ground : 124.3
For the El Salvador washed coffee, the AI did a solid job. The variation between the lightest and darkest ground coffee samples was only 7 Agtron points, and the deviation from the average was just 4 points. That’s a respectable level of consistency!
The Ethiopian natural coffee, however, was another story. It was like trying to get a classroom of hyperactive kids to sit still—pure chaos. The difference between the lightest and darkest roasts was a staggering 39.5 Agtron points, and even the deviation from the average was 23 points. The biggest outlier? Roast No. 4 (7BD), which hit an extreme 126.6 Agtron. That’s roast inconsistency on another level.
The two most extreme Ethiopian roasts—7BD and 42F—had wildly different first crack timings:
7BD: First crack at 4:44 min, 196°C
42F: First crack at 6:07 min, 207°C
That’s an insane 1-minute 23-second difference in FC timing. It’s like one coffee sprinted to the finish line, while the other took a casual stroll and still beat its friend to the end. If you look at moisture release graph you can see that roast 42F had a much more intensive moisture release compared to 7BD hence Ikawa had an easier time to detect it.
Moisture Release Variability: Ethiopian naturals, especially those with mixed bean sizes, don’t release moisture in a uniform way. Some beans in the batch may have released moisture earlier than others, leading the AI to detect an early peak. Take a look at the size distribution of the Ethiopian beans—it’s like a coffee bean family reunion where giants and minis are sharing the same space, each bringing their own roasting quirks to the mix.
Density & Bean Size Variation: This natural coffee had a wide range of bean sizes and densities, which impacted heat absorption. Smaller, less dense beans might have hit first crack earlier, while larger, denser beans needed more time to develop.
“The Ethiopian natural coffee, however, was another story. It was like trying to get a classroom of hyperactive kids to sit still—pure chaos.”
This experiment was a small-scale test, and to truly evaluate AI first crack detection, a much larger study would be needed. More coffees, more roasts, and a wider variety of processing methods, moisture contents, densities, and size distributions would give a fuller picture. Believe it or not, even this simple experiment took multiple days of dedication—scaling it up would be a serious project.
That said, based on these results, the AI feature has its limits. It doesn’t seem to handle coffees with non-uniform moisture release well, like the Ethiopian natural in this test. But would all natural coffees behave the same way? I doubt it. I’ve had great success roasting high-end natural Geishas using AI FC detection, and they came out beautifully. Other times, though, they’ve been underdeveloped—but that was rare. And honestly, that’s what led me to do this experiment in the first place.
Would I consider the AI feature useful? 100%! It’s a powerful tool, but like any tool, it needs to be used with a bit of common sense and human judgment. IKAWA is pioneering something exciting here, and with further fine-tuning, this technology has the potential to redefine coffee roasting as we know it.
But wait—there’s more! I ran the same experiment using profile replication instead of AI first crack detection. What do you think happened? Did it outperform AI? Did it struggle just as much with the Ethiopian natural? That’s a story for next time. 😉 Stay tuned
Author :
TM CoffeeIs.Me
Copyright © 2025 All rights reserved.
Session expired
Please log in again. The login page will open in a new tab. After logging in you can close it and return to this page.
