Free-scale quantification of biomass and fucoxanthin in mass culture of Phaeodactylum tricornutum using Raman spectroscopy coupled with machine learning and transfer learning

Rapid, non-invasive monitoring technologies represent an advancement beyond labor-intensive assays for assessing biomass and bioactives in microalgal production. This study established an integrated Raman spectroscopy and machine learning framework for quantification of <i>Phaeodactylum tricornutum</i> biomass and fucoxanthin yield. Workflow combined spectral processing with multiple feature selection strategies and machine learning algorithms. Deep networks demonstrated superior biomass prediction accuracy on full-spectrum (coefficient of determination R² = 0.968; root mean squared error RMSE = 0.045 g L^-1; relative prediction deviation RPD = 5.598), while kernel methods achieved robust fucoxanthin quantification through sparse, regularized features (R² = 0.949, RMSE = 1.112 mg L^-1, RPD = 4.408). Direct application of laboratory-trained models to pilot-scale photobioreactors was inaccurate due to domain shifts. Transfer learning effectively mitigated these shifts, restoring accuracy with minimal calibration efforts (10-20%). This work presents an industrially viable analytical tool for quality control and process optimization of microalgae-derived bioproducts.