AI Insight
Researchers developed a hybrid machine learning model trained on 963 fermentation experiments to optimize yeast protein production from food waste. The model uses transfer learning across 46 yeast strains and 79 food waste substrates, combined with Bayesian optimization to efficiently identify optimal fermentation conditions. When tested on five independent datasets, this approach reduced the number of required experimental trials by an average of 66% compared to traditional optimization methods.
Why it matters
This work addresses food security and sustainability by making it more efficient to convert diverse food waste streams into protein for human consumption or animal feed. The computational approach could significantly accelerate the development of circular economy solutions by reducing the time and cost of optimizing bioprocesses for variable waste substrates.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Food production is a significant contributor to global greenhouse gas emissions and deforestation, exacerbated by substantial food waste. Converting food waste into yeast protein offers a sustainable solution to enhance food security and contribute to a circular economy. However, due to the diverse and variable nature of food waste substrates, numerous experimental trials are required to optimise the preprocessing steps, yeast strain selection, nutrient addition, and fermentation conditions. This study presents a hybrid modelling approach where data-driven machine learning is used to predict microbial growth kinetics from process parameters. The hybrid model was trained on a comprehensive dataset consisting of 963 fermentation experiments from 55 publications, enabling transfer learning across 46 yeast strains and 79 food waste substrates. The hybrid modelling method was integrated with Bayesian optimisation, a sequential strategy to optimise expensive-to-evaluate functions, to efficiently maximise yeast biomass growth from different food waste substrates. The utility of the hybrid model was evaluated using five test datasets selected from previous literature and was shown to facilitate an average reduction of 66% in the number of experimental trials required to identify optimal fermentation conditions compared to without using the hybrid model. This proved that the transfer of knowledge between yeast strains and food wastes improved the optimisation efficiency of real, previously published datasets compared to traditional optimisation methods. The novelty and contributions of this study include the collation of the extensive dataset, provided as supplementary material; and the demonstration that transfer learning by training the hybrid model on this heterogeneous dataset can improve the optimisation efficiency for yeast biomass growth on new strains and substrates.