Flavor-Cyber-Agriculture: Optimization of plant metabolites in an open-source control environment through surrogate modeling

Arielle J. Johnson, Elliot Meyerson, John de la Parra, Timothy Savas, Risto Miikkulainen, Caleb HarperbioRxiv 2018 · DOI

TL;DR

This paper introduces a machine learning-driven approach to optimize flavor compounds in basil grown in a controlled environment (cyber-agriculture). By using symbolic regression to build surrogate models of plant responses to environmental conditions, they efficiently discovered novel "climate recipes" that significantly boosted flavor molecules, even revealing that 24-hour light cycles are optimal for basil flavor.

What Problem Does This Paper Address?

Conventional agriculture faces challenges in meeting demand, reducing waste, and maintaining food quality, particularly flavor and nutrition. Cyber-agriculture, or growing plants in controlled environments, offers a potential solution by allowing precise manipulation of environmental factors to achieve desired plant outcomes. However, optimizing these complex "climate recipes" for specific traits like flavor is difficult and time-consuming using traditional experimental methods. This paper addresses the challenge of efficiently discovering optimal environmental conditions for maximizing flavor in basil.

Methods

The core of the methodology involves a surrogate-based optimization approach.

  • Cyber-agriculture setup: A hydroponic farm, termed the "Food Computer," was used to control environmental variables.
  • Environmental variables (Design Variables): Photoperiod (hours of light per day), UV light period, and Photosynthetically Active Radiation (PAR) were manipulated.
  • Flavor Measurement: Gas Chromatography-Mass Spectrometry (GC-MS) was used to measure flavor-active volatile molecules.
  • Target Metric: A "chemscore" was developed as a weighted average of volatile profiles compared to control conditions, serving as a holistic measure of flavor.
  • Surrogate Modeling: Symbolic regression was employed to build predictive models (surrogate models) that map environmental variables to the chemscore. This method discovers interpretable algebraic expressions.
  • Optimization: A multi-objective Pareto optimization procedure was used to balance model accuracy (minimizing error) and simplicity (minimizing expression size). The surrogate models were then used to search for new, optimized "climate recipes."
  • Iterative Process: The discovered recipes were tested in the Food Computer, and the results fed back into the surrogate model for further refinement over three rounds of experiments.

Datasets

The study utilized data generated from experiments conducted within the OpenAg Food Computer. This involved growing basil plants under various controlled environmental conditions and measuring their volatile compounds. The data collected from each experimental round served as the dataset for training and refining the surrogate models for subsequent rounds.

Key Insights and Ideas

  • 24-Hour Light is King for Basil Flavor: The most surprising discovery was that a continuous 24-hour photoperiod (all-day light) significantly boosted flavor molecule production in basil, contradicting the initial assumption that a dark period might be necessary.
  • The "Dilution Effect" Confirmed: The study demonstrated a negative correlation between plant weight and flavor compounds, reinforcing the known "dilution effect" where increased biomass can lead to decreased concentration of desirable compounds.
  • Surrogate Optimization for Discovery: Machine learning, specifically symbolic regression, proved highly effective in discovering novel and surprising "climate recipes" that outperformed hand-designed ones, significantly reducing the time and effort required for optimization.
  • Photoperiod Dominance: Analysis revealed that photoperiod was the most dominant environmental variable influencing flavor, often overshadowing the effects of UV light and PAR in the models.

Key Results

  • Several newly discovered "climate recipes" resulted in a marked increase in volatiles compared to control conditions.
  • In the third round, many conditions with extended PAR photoperiods (22.5-24 hours) and UV periods (4-17 hours) doubled the volatile profile compared to control.
  • The surrogate optimization approach successfully iterated towards optimal recipes, with the 24-hour photoperiod emerging as a key factor for maximizing flavor.
  • Spearman correlation analysis showed a strong positive correlation between R-Score (a measure of overall volatile increase) and photoperiod (0.763), and a negative correlation with weight (-0.355), highlighting the flavor-weight tradeoff.

Limitations and Future Work

  • Limited Search Space: The initial experiment focused on a relatively small set of design variables (photoperiod, UV, PAR) and a limited number of iterations.
  • Future Work:
    • Expand Search Space: Develop a larger facility with more actuators (temperature, pH, nutrients, etc.) to explore a wider range of environmental conditions and plant cultivars.
  • Multiple Objectives: Incorporate other objectives beyond flavor, such as nutrient content, density, and yield.
  • Time-Varying Recipes: Move from static recipes to dynamic, real-time optimized recipes that adapt to the plant's growth stage and environmental signals.
  • Other Plants: Extend the methodology to optimize other crops, potentially for different traits (e.g., fiber strength in cotton).
  • Advanced Models: Explore more complex machine learning models (e.g., deep neural networks) for higher-dimensional problems and neuroevolution for real-time control.

Summary updated

Abstract

Abstract Food production in conventional agriculture faces numerous challenges such as reducing waste, meeting demand, maintaining flavor, and providing nutrition. Contained environments under artificial climate control, or cyber-agriculture, could in principle be used to meet many of these challenges. Through such environments, phenotypic expression of the plant---mass, edible yield, flavor, and nutrients---can be actuated through a “climate recipe,” where light, water, nutrients, temperature, and other climate and ecological variables are optimized to achieve a desired result. This paper describes a method for doing this optimization for the desired result of flavor by combining cyber-agriculture, metabolomic phenotype (chemotype) measurements, and machine learning. In a pilot experiment, (1) environmental conditions, i.e. photoperiod and ultraviolet (UV) light (known to affect production of flavor-active molecules in edible plants) were applied under different regimes to basil plants ( Ocimum basilicum ) growing inside a hydroponic farm with an open-source design; (2) flavor-active volatile molecules were measured in each plant using gas chromatography-mass spectrometry (GC-MS); and (3) symbolic regression was used to construct a surrogate model of this chemistry from the input environmental variables, and this model was used to discover new combinations of photoperiod and UV light to increase this chemistry. These new combinations, or climate recipes, were then implemented in the hydroponic farm, and several of them resulted in a marked increase in volatiles over control. The process also led to two important insights: it demonstrated a “dilution effect”, i.e. a negative correlation between weight and desirable chemical species, and it discovered the surprising effect that a 24-hour photoperiod of photosynthetic-active radiation, the equivalent of all-day light, induces the most flavor molecule production in basil. In this manner, surrogate optimization through machine learning can be used to discover effective recipes for cyber-agriculture that would be difficult and time-consuming to find using hand-designed experiments.
BibTeX
@misc{johnson2018flavor,
  author = {Arielle J. Johnson and Elliot Meyerson and John de la Parra and Timothy Savas and Risto Miikkulainen and Caleb Harper},
  title = {{Flavor-Cyber-Agriculture: Optimization of plant metabolites in an open-source control environment through surrogate modeling}},
  year = {2018},
  doi = {10.1101/424226},
  howpublished = {bioRxiv},
}
Figure