Integration of spectroscopic techniques and machine learning for optimizing Phaeodactylum tricornutum cell and fucoxanthin productivity

Abstract

The development of sustainable and controlled microalgae bioprocesses relies on robust and rapid monitoring tools that facilitate continuous process optimization, ensuring high productivity and minimizing response times.

In this work, we analyse the influence of medium formulation on the growth and productivity of axenic Phaeodactylum tricornutum cultures and use the resulting data to develop machine learning (ML) models based on spectroscopy. Our culture assays produced a comprehensive dataset of 255 observations, enabling us to train 55 (24+31) robust models that predict cells or fucoxanthin directly from either absorbance or 2D-fluorescence spectroscopy.

We demonstrate that medium formulation significantly affects cell and fucoxanthin concentrations, and that these effects can be effectively monitored using the developed models, free of overfitting. On a separate data subset, the models demonstrated high accuracy (cell: R² = 0.98, RMSEP = 2.41x10⁶ cells/mL; fucoxanthin: R² = 0.91 and RMSEP = 0.65 ppm), providing a practical, cost-effective, and environmentally friendly alternative to standard analytical methods.