“If you’re working with cereals, for example, it’s a lot more predictable, but with protein, even slight differences in ingredients or any of the parameters can make a huge difference in the end product.”
Tweak any of the parameters, from temperature to pressure, moisture content, feed rates, screw speed, cooling die design, and ingredients, and the results could vary dramatically, says Weiss, an AI and machine learning consultant who has just launched GreenProtein AI, a nonprofit seeking to shine a light into the ‘black box’ of plant protein extrusion.
To reach an optimal result with the perfect combination of the variables above by trial and error would take thousands of tests, something few plant-based meat startups can afford when working with an extrusion provider, she says.
Hence the need for some machine learning, says Weiss (NW), who caught up with AgFunderNews (AFN) to discuss the genesis of GreenProtein AI, and how artificial intelligence and machine learning could be more broadly deployed in the alternative protein industry.
AFN: Can you share the origins story for GreenProtein AI?
NW: Studying computational cognition during my bachelor’s degree [in psychology and cognitive science] was my first encounter with machine learning. This was back when we were focusing on how we think our brains work. I spent some time at PayPal working on data science, but since getting my Master’s I’ve been really focusing on AI technology and applying it to places where it might not be obvious.
I’m a vegan and a big believer in shifting our food system away from animals, so GreenProtein AI [which is funded by the nonprofit Food System Innovations] started with me asking myself the question: How can I use my skill set to help boost the [alternative protein] field?
So I started doing some research and talking to a lot of people in plant-based meat, cultivated meat, and precision fermentation, and mapped out those areas where data science could be used to solve problems that a lot of different companies share.
And that’s where I landed on extrusion, which is an amazing technology with high capacity and an established infrastructure, but where there are also some real pain points.
AFN: Why focus on extrusion for plant-based meat?
NW: It’s something that I’ve heard over and over again, every time I talk to someone from the plant-based meat industry. What they tell me is that extrusion is so unpredictable, that it’s really more of an art than a science, and that trials are so expensive, especially for the smaller food startups. They can’t afford to run all the trials that they need to get to the texture that they really want.
AFN: Why is it so unpredictable?
NW: So maybe this [pea or soy protein] crop had more sunlight when it was grown than the previous one? Maybe the manufacturer changed the extraction methods and the protein powder is slightly different in this batch? Before you even get to changing all the parameters in the machine, the ingredients make a big difference.
AFN: How can machine learning optimize the extrusion process?
NW: Machine learning has the advantage of being able to learn once it’s given enough data; that’s the key. Once it has that data, it can learn from patterns within the data, and then make its predictions. So the goal here is a product based on AI models that researchers and startups could use as a simulation.
So you put in whatever ingredients you want to use, and whatever parameters you are thinking of using, and then you get predictions of what texture you will get. Then you can use that before spending money on extrusion trials.
AFN: Since you can’t eat what comes out of a simulation, how do you assess what the AI model is generating?
NW: There are [objective] ways to measure and describe texture, so there is data from texture analyzers. So someone running the simulation could see, if I use those ingredients and these parameters, that’s the result I get for chewiness, how easily it falls apart, and so on.
But for the second iteration of our product, what we hope to do is to be able compare what emerges from your simulation with the [target] texture you hope to achieve. So if you’re making chicken nuggets, we’ll have the measurements [desired textural characteristics] for your nuggets and we’ll be able to direct you to where you need to go [by adjusting the variables in the simulation] to achieve that.
AFN: But where is the data that will feed into these simulations coming from?
NW: We are right now at the stage of recruiting our first round of seed collaborators, research institutes that have extrusion data and are willing to share it. In addition to being able to use our product, they will get access to the pool of data, which is anonymized, so no IP is compromised.
A lot of commercial companies are very careful about sharing their data, and I get that. Even if someone tells you that everything is protected, I get why people are cautious. So for these companies we will have a second tier of being able to use our model without data sharing.
AFN: Are you the first to apply AI and ML to help optimize extrusion?
NW: There have been attempts, but they were all in-house or specific projects where one lab got an extruder and then just worked on that data. The problem with that is that you have a very small data set, and if you’re just using data from your own extrusion trials, it’s less valuable.
This [GreenProtein AI’s work] will be the first attempt that I know of at pulling data from multiple different players in the field.
AFN: Where do you see potential for AI and machine learning elsewhere in the food industry?
NW: What we are doing with extrusion is the first step of creating a data pool, but it’s something that I would really like to see happening in other areas as well. So cultivated meat is very similar to extrusion in the sense that every company has its own data.
It’s not a lot of data on its own, but if third party were to pull all of that data, it could be very powerful.
I am part of the Good Food Institute (GFI) mentor program and every time I speak to a cultivated meat startup they ask how can they use AI to solve their problems? And my answer is that right now, you probably can’t.
Usually what I say is make sure you collect all of your data, so everything is saved such that you could use it later. If there were a third party neutral player that could aggregate everything in a way such that everyone’s IP were protected, that would be very powerful.
AFN: When it comes to data, we often hear the phrase, ‘Garbage in, garbage out.’ Is that a problem for the deployment of machine learning?
NW: I remember a few years ago a meme was going around along the lines of, ‘What my friends think I do, what my parents think I do, and what I actually do. So what people think that data scientists and machine learning engineers such as myself do, is develop all these amazing new algorithms.
What we actually do for a large percentage of our time is aggregate data, clean data, and make sure everything is standardized. And that will be work that needs to be done [across the industry].
AFN: Do you see potential for machine learning in precision fermentation for optimizing microbial strains and fermentation processes?
NW: Yes, and here the big advantage is that genome data exists and we can use it, which is probably why you see it used more here, whereas for extrusion and cultivated meat, the data isn’t there yet.
What is extrusion?
The most widely-used process to texturize plant proteins, extrusion is a mechanical process whereby plant proteins, water or/and oil and dry ingredients are exposed to heat, moisture and pressure in a chamber with screws that convey material toward a die that provides the final shape to the product.
During the process, the proteins undergo a series of structural modifications, ranging from denaturation to unfolding, crosslinking, and alignment, resulting in a fibrous structure that mimics animal muscle tissue.
Extrusion can be performed at a low moisture level (<30%) to make texturized vegetable protein (TVP), which can be stored at room temperature; or at a high moisture level (>50%) to create more meaty analogs that require refrigeration.