Open Agriculture and Artificial Intelligence for Natural Flavor

by Arielle Johnson, PhD and the Open Agriculture Initiative Team

When you think about it, artificial intelligence has infiltrated almost every aspect of our lives: from cell phones to web searches, the advertisements we watch and the products we buy. AI is everywhere, including in our food system.

At the 2015 Techonomy conference in Half Moon Bay, Media Lab’s Caleb Harper laid out his vision for the Open Agriculture Initiative (OpenAg), and our group’s mission to give anyone access to the open source tools and experimental data necessary to grow food, wherever they are (visit that conference here). Babak Hodjat, the cofounder of Sentient, was sitting in the audience and recognized that his company’s deep machine learning and AI platform could make sense of all the data that OpenAg’s Food Computers were generating — and then learn from and adapt it. It had already been tested by hospitals looking to predict infections and by traders hoping to determine patterns in the market. Why not for food and agriculture, too?

In the agriculture industry today, there is no open database of “optimal” growing conditions with correlated outcomes for cultivating food. Historically, farmers relied on their own instinct, oral histories, and expertise to come up with specific growing conditions and nutrients for the right “recipe”. The goal was always the same — obtain the best crop outcomes for yield, flavor, and nutrition (to name a few).

But you can’t have the perfect recipe when the variables keep changing, and that means each year and every crop are a gamble. Adding another layer of complexity, many crops only grow in certain climates and are expensive to transport. Both growing and getting certain foods in certain places in the world is simply too resource-intensive.

So what if we could grow crops in controlled environments that maintain the perfect recipe of growing conditions at all times, and learn how to improve upon that recipe over time? If food can be grown anywhere and you could use AI to forecast the best yield (or flavor, or nutrient density) with far more certainty, what would that do to the food and agriculture industry?

In the fall of 2015, OpenAg had completed its prototypes of open-source, small and medium-scale hydroponic controlled-environment growth chambers that we call the Food Computers. Equipped with sensors, actuators and the open-source microcontrollers Arduino and Raspberry Pi, Food Computers are a great experimental tool because they allows us to monitor and adapt the environment of the plants growing inside, and then analyze which combinations of environmental or biochemical variables correlate to the best phenotypic response, or outcome.

We think flavor is one of the most important tools we have for deciding how to eat well, and as a flavor chemist and OpenAg’s Visiting Scientist, I suggested that our first experiments in Food Computers be to test optimization for flavor. We chose basil (Ocimum basilicum) as a model organism, building on research that shows that flavor molecule concentrations in basil increase after exposure to stress (e.g. increased heat, salinity, UV light, water stress, chitosan).

But what exactly is a climate recipe for flavor, and how does a Food Computer help us optimize it? Imagine a group of expert farmers each come up with their ideal set of environmental conditions for growing very flavorful basil — light, soil, water, climate, etc. We translate those conditions into actuatable code, run that code in a Food Computer, and test and correlate the levels of flavor (in this case, volatile molecules called monoterpenes, sesquiterpenes, and phenylpropenes) generated by each plant, each time its is grown. The results from those tests inform the next round of hypotheses for what combinations of environmental conditions produce the most flavorful basil, and the process continues, evolving into an optimized climate recipe as the generations proceed, and without the limit of conventional seasons. The optimized climate recipe for flavorful basil may resemble the climate is Genoa, Italy (home of the famed Genovese Basil) or in its iterative development, it may become a climate found nowhere on Earth.

Food Computers generate roughly three-million points of data, per plant, per growth cycle — that’s a lot of data to comb through and analyze. We’re keeping this data open and accessible to the public so that anyone in the community can become part of our research, and test and improve climate recipes over time. In an ideal world, Food Computers will learn from the data as they’re generating it and the climate recipes will get better through the growth cycle.

We’ve run our first three rounds of basil experiments’ data through Sentient’s AI and we’ve already learned a few things. First, our model discovered a strong negative correlation between weight and flavor — the bigger the basil plant, the less concentrated the flavor. This is actually a well-known phenomenon in agriculture (resulting from competition between primary and secondary metabolism) but it was very encouraging to see the model pick this up so quickly.

Next, the model also discovered that flavor improvements were achieved with constant, 24-hour light periods. This surprised our team who thought basil needed a natural rest cycle. And lastly, our results showed a number of significant, nonlinear interactions between recipe variables. This confirms that optimization using surrogate modeling is a good fit for this problem. In short, the model is finding things we would never have been able to find on our own.

In the future, the methodology can be extended to optimize other outcomes, such as nutrient density or taste. It is possible to include other plants, such as cotton, to optimize fiber quality (length, strength, and fineness). It is also possible to make optimization multiobjective, and include for example, cost (kWh etc.) as well. Applications such as bioengineering, biofuels, drug design, and more, are likewise possible.

We’ve been on a bit of a hiatus with basil experiments as we build our new Food Server lab at our new facility in Middleton, Massachusetts, where we’ll continue our flavor-hacking experiments as we bring more controlled environments online. To get your hands dirty with our data, it is all available, open and for free, on our Github. If you are a plant scientist or analytical chemist, or just someone who really wants to run an analysis on some of our plant tissue samples (strictly to be released as open data!), drop us a line.

Thanks to Sentient for sharing the results of their machine learning analysis with us, and to Ginkgo Bioworks for letting us use their GC-MS on the weekends for our volatile analysis.