Predictive modeling of Ulva sp. growth and chemical composition in an outdoor air-mixed bioreactor under natural environmental conditions: A machine learning approach

Abstract

Approximately 35 million tons of wet macroalgae were harvested from aquaculture in 2020, and its cultivation is rapidly increasing. However, annual variability in yield and chemical composition due to natural environmental conditions makes cost-benefit analysis difficult, hindering profitable macroalgae cultivation.

This study aims to develop models for predicting the growth and chemical composition of the green seaweed Ulva sp. based on measurable environmental variables. We used Forward Selection Search (FSS), the Ordinary Least Squares (OLS) best subset approach, and LASSO to develop a prediction model from two years of experimental measurements of Ulva sp. biomass growth and chemical composition in Mikhmoret, Israel. The best predictive model for fresh mass achieved an R² of 0.77 with a Mean Absolute Percentage Error (MAPE) of 32 %. For dry mass, the R² was 0.75 with a significantly higher MAPE of 62 %. The prediction for ash-free dry mass yielded an R² of 0.6 and a Root Mean Square Error (RMSE) of 0.62. Carbon content prediction attained an R² of 0.70 with an RMSE of 0.49, while nitrogen content prediction resulted in an R² of 0.69 with an RMSE of 0.56.

Our study demonstrates the potential of using machine learning to analyze seagricultural data and understand the yield and chemical composition in Ulva sp. These results could lead to the development of optimized cultivation techniques for large-scale seaweed farming.