Software suggestion — A.I. Profile Design

Have you discovered a good way to do something with your Kaffelogic? Share it here.
Post Reply
benjaminfleon
Posts: 29
Joined: Wed 28 Apr, 2021 5:02 am
Location: Santiago, Chile
x 33

Software suggestion — A.I. Profile Design

#1983

Post by benjaminfleon »

Hello everyone!

I've been having a lot of conversations with a few friends and colleagues about the idea of a program that could design profile curves based on machine learning / AI.

The basic idea is that with enough information, we could teach an artificial intelligence to design profiles based on a desired sensory outcome; which could potentially also be applied to other roasting machines that at least partially rely on automation.

To do this, we're thinking of 3 main areas that would need to be described or factored in the equation, namely:
— Green coffee and their most significant characteristics (these may be density, moisture content, size, etc; we have a rough idea of which are the most important, but I haven't encountered anything definitive in scientific literature anyways).
— Roast curve parameters for a given batch size (end temp, total time, build-up*/maillard/development duration and proportion, RoR, AUC, etc).
— Sensory evaluation (acidity, sweetness, bitterness, body, clarity, complexity, balance, etc).

* I personally call the "drying phase" this because I think it's a little bit more appropriate, but you're welcome to disagree, it's just a personal preference.

We realize this would take a lot of careful data collection, which is what brings me here.

I wonder if its possible to — or if its an idea worth exploring in your opinion—, to
1. Include the area under the curve as an option for the phases panel in order to make data collection easier (and sort of figure out if this data point is significant or not in the long run). This would mean that the AUC of each phase (with the temp at the start of the phase as the base temp) would appear under the temp increase; and that the total AUC (with the starting temp as a base temp) would appear under the total roast time.
2. Include the feature to go from one log to the previous and next one by pressing a button somewhere in the window, without the need to go to the "view memory stick" button.
3. Include visualization of RoR and power curves from different logs when comparing them.
4. Include the option to change the color of different curves, both within one log, and when comparing several.

I realize it's a lot to ask all at once, but this definitely is a long term project and I don't lose anything by asking.

I hope everybody is doing well, thank you in advance!
benjaminfleon
Posts: 29
Joined: Wed 28 Apr, 2021 5:02 am
Location: Santiago, Chile
x 33

Re: Software suggestion — Profile stretching

#2021

Post by benjaminfleon »

Hello to whoever is reading this!

I've recently published a bunch of profiles in the user contributed profiles section, and I've been thinking about the possibility of including a feature for moving several or all the points of the curve at once, as well as stretching or shortening it in time.

That would let us try a profile that works particularly well in different lengths of time with the same relative proportions of the different phases, fan speed profile and end temperature.

After stretching a particular profile, we would need to adjust the preheat power, which Studio already recommends, so I think this could be potentially super useful for a lot of users.
User avatar
kaffelogic
Site Admin
Posts: 498
Joined: Thu 18 Oct, 2018 9:38 am
x 72
Contact:

Re: Software suggestion — A.I. Profile Design

#2022

Post by kaffelogic »

benjaminfleon wrote: Mon 31 May, 2021 7:24 am Hello everyone!

I've been having a lot of conversations with a few friends and colleagues about the idea of a program that could design profile curves based on machine learning / AI.

The basic idea is that with enough information, we could teach an artificial intelligence to design profiles based on a desired sensory outcome; which could potentially also be applied to other roasting machines that at least partially rely on automation.

To do this, we're thinking of 3 main areas that would need to be described or factored in the equation, namely:
— Green coffee and their most significant characteristics (these may be density, moisture content, size, etc; we have a rough idea of which are the most important, but I haven't encountered anything definitive in scientific literature anyways).
— Roast curve parameters for a given batch size (end temp, total time, build-up*/maillard/development duration and proportion, RoR, AUC, etc).
— Sensory evaluation (acidity, sweetness, bitterness, body, clarity, complexity, balance, etc).

* I personally call the "drying phase" this because I think it's a little bit more appropriate, but you're welcome to disagree, it's just a personal preference.

We realize this would take a lot of careful data collection, which is what brings me here.

I wonder if its possible to — or if its an idea worth exploring in your opinion—, to
1. Include the area under the curve as an option for the phases panel in order to make data collection easier (and sort of figure out if this data point is significant or not in the long run). This would mean that the AUC of each phase (with the temp at the start of the phase as the base temp) would appear under the temp increase; and that the total AUC (with the starting temp as a base temp) would appear under the total roast time.
2. Include the feature to go from one log to the previous and next one by pressing a button somewhere in the window, without the need to go to the "view memory stick" button.
3. Include visualization of RoR and power curves from different logs when comparing them.
4. Include the option to change the color of different curves, both within one log, and when comparing several.

I realize it's a lot to ask all at once, but this definitely is a long term project and I don't lose anything by asking.

I hope everybody is doing well, thank you in advance!
I agree with what you are saying about AI. This big challenge with training a neutral network to design profiles is 1. efficiently obtaining the objective data that describes the green coffee and 2. measuring the sensory outcome in a way that correlates with what people actually like. This is not too say it can't be done, but it is a significant body of work to do it.

As for your requested UI improvements to Studio, I have added them to our backlog of UI improvements with a view to making those improvements when time and resourcing permits. So no promises as to when, but thank you for the suggestions and we will act on them as they are all worthy improvements.
User avatar
kaffelogic
Site Admin
Posts: 498
Joined: Thu 18 Oct, 2018 9:38 am
x 72
Contact:

Re: Software suggestion — Profile stretching

#2023

Post by kaffelogic »

benjaminfleon wrote: Fri 09 Jul, 2021 11:47 pm Hello to whoever is reading this!

I've recently published a bunch of profiles in the user contributed profiles section, and I've been thinking about the possibility of including a feature for moving several or all the points of the curve at once, as well as stretching or shortening it in time.

That would let us try a profile that works particularly well in different lengths of time with the same relative proportions of the different phases, fan speed profile and end temperature.

After stretching a particular profile, we would need to adjust the preheat power, which Studio already recommends, so I think this could be potentially super useful for a lot of users.
The ability to stretch an entire profile is available with the transformation tool (tools menu) of Studio. It's not exactly well documented, but it's especially used by people transferring profiles from the Kaffelogic to, for example, a production drum roaster.

The ability to transform just selected points has not been implemented, but I have added it to our list of suggested improvements.
Post Reply